Search

Aerosol research is important for the protection of the ecological environment, the improvement of air quality, and as a response to climate change. In this study, a random forest (RF) estimation model of aerosol optical depth (AOD) and extinction coefficient vertical profiles was, respectively, established using Orbiting Carbon Observatory-2 (OCO-2) oxygen-A band (O2 A-band) data from China and its surrounding areas in 2016, combined with geographical information (longitude, latitude, and elevation) and viewing angle data. To address the high number of OCO-2 O2 A-band channels, principal component analysis (PCA) was employed for dimensionality reduction. The model was then applied to estimate the aerosol extinction coefficients for the region in 2017, and its validity was verified by comparing the estimated values with the Cloud-Aerosol Lidar Infrared Pathfinder Satellite Observations (CALIPSO) Level 2 extinction coefficients. In the comprehensive analysis of overall performance, an AOD model was initially constructed using variables, achieving a correlation coefficient (R) of 0.676. Subsequently, predictions for aerosol extinction coefficients were generated, revealing a satisfactory agreement between the predicted and the actual values in the vertical direction, with an R of 0.535 and a root mean square error (RMSE) of 0.107 km−1. Of the four seasons of the year, the model performs best in autumn (R = 0.557), while its performance was relatively lower in summer (R = 0.442). Height had a significant effect on the model, with both R and RMSE decreasing as height increased. Furthermore, the accuracy of aerosol profile inversion shows a dependence on AOD, with a better accuracy when AOD is less than 0.3 and RMSE can be less than 0.06 km−1. [ABSTRACT FROM AUTHOR]

Moso bamboo (Phyllostachys edulis) is a crucial species among the 500 varieties of bamboo found in China and plays an important role in providing ecosystem services. However, remote sensing studies on the invasion of Moso bamboo, especially its impact on forest biodiversity, are limited. Therefore, we explored the feasibility of using Sentinel-2 multispectral data and digital elevation data from the Shuttle Radar Topography Mission and random forest (RF) algorithms to monitor changes in forest diversity due to the spread of Moso bamboo. From October to November 2019, researchers conducted field surveys on 100 subtropical forest plots in Zhejiang Province, China. Four biodiversity indices (Margalef, Shannon, Simpson, and Pielou) were calculated from the survey data. Subsequently, after completing 100 epochs of training and testing, we developed the RF prediction model and assessed its performance using three key metrics: coefficient of determination, root mean squared error, and mean absolute error. Our results showed that the RF model has a strong predictive ability for all indices except for the Pilou index, which has an average predictive ability. These results demonstrate the feasibility of using remote sensing to monitor forest diversity changes caused by the spreading of Moso bamboo. [ABSTRACT FROM AUTHOR]

Software effort estimation has constituted a significant research theme in recent years. The more important provocation for project managers concerns reaching their targets within the fixed time boundary. Machine learning strategies can lead software management to an entire novel stage. The purpose of this research work is to compare an optimized long short-term memory neural network, based on particle swarm optimization, with six machine learning methods used to predict software development effort: K-nearest neighbours, decision tree, random forest, gradient boosted tree, multilayer perceptron, and long short-term memory. The process of effort estimation uses five datasets: China and Desharnais, for which outputs are expressed in person-hours; and Albrecht, Kemerer, and Cocomo81, for which outputs are measured in person-months. To compare the accuracy of these intelligent methods four metrics were used: mean absolute error, median absolute error, root mean square error, and coefficient of determination. For all five datasets, based on metric values, it was concluded that the proposed optimized long short-term memory intelligent method predicts more accurately the effort required to develop a software product. Python 3.8.12 programming language was used in conjunction with the TensorFlow 2.10.0, Keras 2.10.0, and SKlearn 1.0.1 to implement these machine learning methods. [ABSTRACT FROM AUTHOR]

Topographic variables derived from digital elevation models (DEM) are the most commonly used predictor in soil organic carbon (SOC) prediction. The predictability of these variates for the spatial heterogenity of SOC is influenced by the size of DEM resolution. This paper aimed to investigate the resolution effect of the topography at multiple DEM resolutions and select the most important predictors and the optimal resolution for SOC prediction in a large-scale and morphologically complex area. The study area, covering 15 600 km2, was located in a mountainous province in China. A total of 3901 soil samples and three machine learning algorithms including random forest, gradient boosted regression tree, and artificial neural network were used to construct the SOC predictive models at resolutions of 30, 90, 150, 210, 270, 330, and 390 m. Topographic factors were derived from the DEM with an original resolution of 30 m and were subsequently resampled to varying resolutions by using the bilinear interpolation algorithm. The importance of each variate in each predictive model was computed for the most important predictor selection, and the determination coefficient (R2) and Root Mean Square Error (RMSE) of each predictive model were computed for the accuracy evaluation. Comparative analysis on point-based topographic representation, the accuracy of predictive models, and the variable importance was conducted among different resolutions to investigate the resolution effect on the SOC prediction and to select the optimal resolution for this study area. The results showed that topography was the key factor influencing the SOC distribution in this large-scale area due to the region's mountainous nature, and the relative contribution of the topography to predict SOC distribution varied with DEM resolutions. Elevation was the most important topographic factor at all DEM resolutions. Accuracies of predictive models built on the three machine learning algorithms were all dependent on DEM resolution, indicating that DEM resolution had an important effect on SOC prediction. Overall, the random forest predictive models outperformed the other two algorithms, and its most accurate result was obtained at the 210-m resolution with an R2 of 0.25 and an RMSE of 4.77 g kg–1. In conclusion, we found that finer resolutions did not necessarily produce more accurate SOC predictions even in a large-scale and morphologically complex area. Therefore, investigating the resolution effect is suggested to select an optimal resolution for the SOC prediction in other regions characterized by similar geomorphological conditions as this study area. [ABSTRACT FROM AUTHOR]

Accurate wind speed prediction is crucial for the safe and efficient utilization of wind resources. However, current single-value deterministic numerical weather prediction methods employed by wind farms do not adequately meet the actual needs of power grid dispatching. In this study, we propose a new hybrid forecasting method for correcting 10 m wind speed predictions made by the Weather Research and Forecasting (WRF) model. Our approach incorporates variational mode decomposition (VMD), principal component analysis (PCA), and five artificial intelligence algorithms: deep belief network (DBN), multilayer perceptron (MLP), random forest (RF), eXtreme gradient boosting (XGBoost), light gradient boosting machine (lightGBM), and the Bayesian optimization algorithm (BOA). We first predict wind speeds using the WRF model, with initial and lateral boundary conditions from the Global Forecast System (GFS). We then perform two sets of experiments with different input factors and apply BOA optimization to tune the four artificial intelligence models, ultimately building the final models. Furthermore, we compare the aforementioned five optimal artificial intelligence models suitable for five provinces in southern China in the wintertime: VMD-PCA-RF in December 2021 and VMD-PCA-lightGBM in January 2022. We find that the VMD-PCA-RF evaluation indices exhibit relative stability over nearly a year: the correlation coefficient (R) is above 0.6, forecasting accuracy (FA) is above 85 %, mean absolute error (MAE) is below 0.6 m s -1 , root mean square error (RMSE) is below 0.8 m s -1 , relative mean absolute error (rMAE) is below 60 %, and relative root mean square error (rRMSE) is below 75 %. Thus, for its promising performance and excellent year-round robustness, we recommend adopting the proposed VMD-PCA-RF method for improved wind speed prediction in models. [ABSTRACT FROM AUTHOR]

Among the critical global crises curbing world development and sustainability, air quality degradation has been a long-lasting and increasingly urgent one and it has been sufficiently proven to pose severe threats to human health and social welfare. A higher level of model prediction accuracy can play a fundamental role in air quality assessment and enhancing human well-being. In this paper, four types of machine learning models—random forest model, ridge regression model, support vector machine model, extremely randomized trees model—were adopted to predict PM2.5 concentration in ten cities in the Jing-Jin-Ji region of north China based on multi-sources spatiotemporal data including air quality and meteorological data in time series. Data were fed into the model by using the rolling prediction method which is proven to improve prediction accuracy in our experiments. Lastly, the comparative experiments show that at the city level, RF and ExtraTrees models have better predictive results with lower mean absolute error (MAE), root mean square error (RMSE), and higher index of agreement (IA) compared to other selected models. For seasonality, level four models all have the best prediction performances in winter time and the worst in summer time, and RF models have the best prediction performance with the IA ranging from 0.93 to 0.98 with an MAE of 5.91 to 11.68 μg/m3. Consequently, the demonstration of how each model performs differently in each city and each season is expected to shed light on environmental policy implications. [ABSTRACT FROM AUTHOR]

The normalized difference vegetation index (NDVI) is one of the most common metrics used to describe vegetation dynamics. Unfortunately, low-quality pixels resulting from contamination (by features including clouds, snow, aerosols, and mixed factors) have impeded NDVI products' widespread application. Researchers have thought of several ways to improve NDVI quality when contamination occurs. However, most of these algorithms are based on the noise-negative deviation principle, which aligns low-value NDVI products to an upper line but ignores cases where absolute surface values are low. Consequently, to fill in these research gaps, in this article, we use the random forest model to produce a set of high-quality NDVI products to represent actual surface characteristics more accurately and naturally. Climate and geographical products are used as model inputs to describe environmental factors. They represent the random forest (RF) model that establishes relationships between MODIS NDVI products and meteorological products in high-quality areas. In addition, auxiliary data and empirical knowledge are employed to meet filling requirements. Notably, the random forest (RF) algorithm exhibits a mean absolute error (MAE) of 0.024 and a root mean squared error (RMSE) of 0.034, in addition to a coefficient of determination (R2) value of 0.974. Furthermore, the MAE and RMSE of the RF-based method decreased by 0.014 and 0.019, respectively, when compared to those of the STSG (spatial–temporal Savitzky–Golay) plan and by 0.013 and 0.015, respectively, when compared to the LSTM (long short-term memory) method. R2 increased by 0.039 and 0.027, respectively, compared to the STSG and LSTM methods. We introduced a novel series of NDVI products that demonstrated consistent spatial and temporal connectivity. The novel product exhibits enhanced adaptability to intricate environmental conditions and promises the potential for utilization in investigating vegetation dynamics within the Chinese region. [ABSTRACT FROM AUTHOR]

Purpose: Integrating process-based models with machine learning (ML) is an important way to derive spatially-explicit information on the dynamics of cropland soil organic carbon (SOC) stock. However, the performance of integrating process-based models with different structures remains unclear. The purpose of this study is to develop integrated models by combining multiple process-based models with ML for improving the spatiotemporal dynamics modelling of cropland SOC. Methods: A total of 1219 cropland SOC stock data (0–20 cm, sampled in 1980, 2000, and 2015) were collected from southern Jiangsu Province of China. Three models were built by integrating either Rothamsted carbon model (RothC), microbial-mineral carbon stabilization model (MIMICS), or the both with space-for-time random forest (RF-SFTS). Results: The overall spatiotemporal patterns of SOC dynamics that modelled by RothC, MIMICS, the integration of RothC with RF-SFTS (RF-RothC), the integration of MIMICS with RF-SFTS (RF-MIMICS), and the integration of RothC and MIMICS with RF-SFTS (RF-RothC-MIMICS) were similar, but their prediction accuracies differed significantly. The RF-RothC-MIMICS model had the best performance among the six models, being a lowest root mean squared error (RMSE) of 0.79 kg C m−2 and a highest coefficient of determination (R2) of 0.54, compared to the prediction accuracy of other models. Conclusions: Integrating process-based models differing in structures with ML significantly improved the spatiotemporal modelling of SOC dynamics. Outputs of this study may provide some guidance on more realistic projection of SOC dynamics under future changes in climate, land use, and management practices. [ABSTRACT FROM AUTHOR]

This study investigated the accuracy of the random forest (RF) model in gap filling the sensible (H) and latent heat (LE) fluxes, by using the observation data collected at a site over rice–wheat rotation croplands in Shouxian County of eastern China from 15 July 2015 to 24 April 2019. Firstly, the variable significance of the machine learning (ML) model's five input variables, including the net radiation (Rn), wind speed (WS), temperature (T), relative humidity (RH), and air pressure (P), was examined, and it was found that Rn accounted for 78 % and 76 % of the total variable significance in H and LE calculating, respectively, showing that it was the most important input variable. Secondly, the RF model's accuracy with the five-variable (Rn , WS, T , RH, P) input combination was evaluated, and the results showed that the RF model could reliably gap fill the H and LE with mean absolute errors (MAEs) of 5.88 and 20.97 W m -2 , and root mean square errors (RMSEs) of 10.67 and 29.46 W m -2 , respectively. Thirdly, four-variable input combinations were tested, and it was found that the best input combination was (Rn , WS, T , P) by removing RH from the input list, and its MAE values of H and LE were reduced by 12.65 % and 7.12 %, respectively. At last, through the Taylor diagram, H and LE gap-filling accuracies of the RF model, the support vector machine (SVM) model, the k nearest-neighbor (KNN) model, and the gradient boosting decision tree (GBDT) model were intercompared, and the statistical metrics showed that RF was the most accurate for both H and LE gap filling, while the LR and KNN model performed the worst for H and LE gap filling, respectively. [ABSTRACT FROM AUTHOR]

Purpose: The present study aimed to reveal the spatial distribution of soil nutrients in southwest China and to quantitatively evaluate the effect of environmental factors on spatial variability of soil nutrients under different cropping systems. Methods: Semivariogram, random forest (RF), and partial dependence plots (PDP) were applied to investigate the relationships between environmental factors and soil nutrients variability. A total of 142 samples were collected from topsoil under rice and maize cropping systems in southwest China. Results: Soil organic carbon (SOC), total nitrogen (TN), and total phosphorus (TP) in the topsoil of the study area were higher than their corresponding background and geochemical baseline values of China. Semivariogram analysis indicated that SOC and TN have weak spatial dependence, and TP has strong spatial dependence. The soil nutrient contents were predicted well by the RF models. The mean absolute error, root mean square error, and coefficient of determination were 0.14–0.47 g/kg, 0.19–3.67 g/kg, and 0.30–0.46, respectively. For different soil nutrients, the relative importance of environmental variables varied greatly. Cropping system, topographic wetness index (TWI), and slope were critical factors that controlled SOC variability. Cropping system, TWI, and mean annual temperature explained most variation of TN. Soil pH, aspect, and cropping system were dominant factors affecting TP variability. Conclusion: In this study, SOC and TN had weak spatial autocorrelation, while TP had strong spatial autocorrelation. Meanwhile, cropping system and topography had greatly impact on the variability of soil nutrients, which should be considered in formulating agricultural measures in the future. [ABSTRACT FROM AUTHOR]

This study investigated the accuracy of the Random Forest (RF) model in gap-filling the sensible (H) and latent heat (LE) fluxes, by using the observation data collected at a site over rice-wheat-rotation croplands in Shouxian County of eastern China from 15 July 2015 to 24 April 2019. Firstly, the variable significances of the machine learning (ML) model's five input variables, including the net radiation (Rn), winds speed (WS), temperature (T), relative humidity (RH), and air pressure (P), were examined, and it was found that Rn accounted for 78 % and 76 % of the total variable significance in H and LE calculating, respectively, showing that it was the most important input variable. Secondly, the RF model's accuracy with the five-variable (Rn, WS, T, RH, P) input combination was evaluated, and the results showed that the RF model could reliably gap-fill the H and LE with mean absolute errors (MAEs) of 5.88 Wm-2 and 20.97 Wm-2, and root mean square errors (RMSEs) of 10.67 Wm-2 and 29.46 Wm-2, respectively. Thirdly, 4-variable input combinations were tested, and it was found that the best input combination was (Rn, WS, T, P) with the MAE of H and LE reduced by 12.65 % and 7.12 %, respectively, after removing RH from the input list. At last, through the Taylor diagram, H and LE gap-filling accuracy of the RF model, the support vector machine (SVM) model, the k-nearest neighbor (KNN) model, and the gradient boosting decision tree (GBDT) model was inter-compared, and the statistical metrics showed that RF was the most accurate for both H and LE gap-filling, while the LR and KNN model performed the worst for H and LE gap-filling, respectively. [ABSTRACT FROM AUTHOR]

Machine learning (ML) has shown high predictive ability in environmental research. Accurate estimation of daily PM2.5 concentrations is a prerequisite to address environmental public health issues. However, studies on the interpretability of ML algorithms were limited. In this study, we aimed to estimate the daily concentrations of PM2.5 at a seasonal level, and to understand the potential mechanisms of ML algorithms' decisions with SHapley Additive exPlanations (SHAP). Daily ground PM2.5 concentrations and meteorological data were obtained from the Beijing Municipal Ecological and Environmental Monitoring Center, and China Meteorological Data Service Centre between December 2013 and 2019 November. We calculated correlation coefficient and variance inflation factor (VIF) to eliminate the variables with collinearity, and recursive feature elimination (RFE) was further used to selected more important predictors. A series of ML algorithms, including linear regression, the variants of linear regression (Ridge, Lasso, Elasticnet), decision tree (DT), k-nearest neighbor (KNN), support vector regression (SVR), ensemble methods (random forest: RF, eXtreme Gradient Boosting: XGBoost), and deep learning (long short-term memory network: LSTM), were developed to estimate seasonal-level daily PM2.5 concentrations. A 10-fold cross validation was used to tune hyperparameters, and root mean square error (RMSE), mean absolute error (MAE), ratio of performance to deviation (RPD), and Lin's concordance correlation coefficient (LCCC) were used to evaluate models' performance. SHAP was performed for local and global interpretability analysis. The results showed that the distribution of PM2.5 concentrations in Beijing showed obvious seasonal patterns. A total of five variables (Precipitation, Mean wind speed, Sunshine duration, Mean surface temperature, Mean relative humidity) were selected for final prediction. LSTM showed much higher accuracy than other traditional ML models, achieved the smallest RMSE of 19.58 µg/m3 and MAE of 15.11 µg/m3. In terms of selected data set, there was acceptable (LCCC = 0.41 ~ 0.52) agreement and accuracy (RPD = 0.97 ~ 1.92) for LSTM. The SHAP analyses revealed that the meteorological factors had different influences in specific predictions, and the complex interactions were also illustrated. These results enhance our understanding of meteorological factors-PM2.5 relationships and explain the mechanisms of ML algorithms' decisions. [ABSTRACT FROM AUTHOR]

In gas drilling operations, the rate of penetration (ROP) parameter has an important influence on drilling costs. Prediction of ROP can optimize the drilling operational parameters and reduce its overall cost. To predict ROP with satisfactory precision, a stacked generalization ensemble model is developed in this paper. Drilling data were collected from a shale gas survey well in Xinjiang, northwestern China. First, Pearson correlation analysis is used for feature selection. Then, a Savitzky-Golay smoothing filter is used to reduce noise in the dataset. In the next stage, we propose a stacked generalization ensemble model that combines six machine learning models: support vector regression (SVR), extremely randomized trees (ET), random forest (RF), gradient boosting machine (GB), light gradient boosting machine (LightGBM) and extreme gradient boosting (XGB). The stacked model generates meta-data from the five models (SVR, ET, RF, GB, LightGBM) to compute ROP predictions using an XGB model. Then, the leave-one-out method is used to verify modeling performance. The performance of the stacked model is better than each single model, with R2 = 0.9568 and root mean square error = 0.4853 m/h achieved on the testing dataset. Hence, the proposed approach will be useful in optimizing gas drilling. Finally, the particle swarm optimization (PSO) algorithm is used to optimize the relevant ROP parameters. [ABSTRACT FROM AUTHOR]

Agricultural drought has been one of the most damaging natural hazards in the world, due mainly to the water shortage. A timely and effective monitoring system can greatly contribute to the management and mitigation of agricultural drought for better crops yields. A drought index can be further used to support the agricultural drought monitoring, assessment, and decision-making on mitigation measures. Therefore, it is a high demand to determine the real drought and monitoring index in practice. Taking the Shandong Province in eastern China as the research area, this study aims to construct a new agricultural drought monitoring model by random forest using the Crop Water Stress Index (CWSI). A deviation correction was also used to construct the Bias-corrected Random Forest Drought Condition Index (BRF-DCI). The physical meaning of evapotranspiration data was elucidated in the occurrence of agricultural drought. The multisource remote sensing was selected, including the Vegetation Condition Index (VCI), Precipitation Condition Index (PCI), Temperature Condition Index (TCI), Soil Moisture Condition Index (SMCI), and Standardized Precipitation Evapotranspiration Index (SPEI). The accuracy of the model was evaluated by the determination coefficient, and root mean square error. Since the study area presents the warm temperate continental monsoon climate with large temporal and spatial changes in the precipitation, some considerations were made on the influence of evapotranspiration on the drought monitoring model, as well as the accuracy and application of drought condition index for different drought grades. The results were as follows: 1) A better performance was achieved, when adding the CWSI as the independent variable into the drought monitoring model, where the determination coefficient of the BRF-DCI index and the observed SPEI-3 was 0.72-0.85, and the root mean square error was 0.58-0.71. 2) The BRF-DCI index with CWSI improved the accuracy of extreme drought monitoring, where the maximum monitoring accuracies of moderate, severe, and special drought were 0.88, 0.89, and 0.91, respectively. As such, the CWSI independent variable significantly improved the accuracy of the model, particularly for the extreme drought monitoring. 3) The drought monitoring index was basically consistent with the drought trend, represented by the real SPEI-3 at different sites, suitable for the changing of the actual drought. 4) The simulation of historical drought using drought condition index was also basically consistent with the actual in the study area. Consequently, the BRF-DCI can be widely expected to accurately predict the drought-affected areas with the temporal and spatial evolution. This finding can provide an important reference to evaluate the agricultural drought monitoring index for the early warning of natural hazards.. [ABSTRACT FROM AUTHOR]

A crop yield estimation can provide the data support for the decision-making in the entire crop growing season. The previous estimation models of winter wheat yield often perform a low accuracy and less generalization, due to the overfitting during machine learning. Meanwhile, the specific feature variables can be expected to reduce the high collinearity between spectral bands of surface reflectance. It is also proper to set the spectral index for the less influence of independent wavelengths on iterative calculation. In this study, an improved estimation model of winter wheat yield was established using Random Forest (RF) in Henan Province of China, using the surface reflectance data and spectral index during the growth period. The specific procedure was listed. 1) The comparisons were conducted for the regional mean values and regional frequency histogram on the spectral bands of surface reflectance. 2) The different spectral bands of surface reflectance were then compared. 3) The generated sample data on the specific feature variable was used to train a yield estimation model. 4) The yield estimation model was applied to evaluate the winter wheat yields during 2013-2015, using the surface reflectance data and spectral index. The estimated yields of winter wheat were also verified by the survey statistics. The results showed that the frequency histogram presented a better performance than the regional mean values, where the Mean Absolute Error (MAE) was 660 kg/hm², Root Mean Square Error (RMSE) was 860 kg/hm², and the coefficient of determination was 0.83. The dimension of input feature variables and computational complexity were greatly reduced to improve the estimation ability of the model. The frequency histogram was suitable for the extraction of feature variables, further to optimize the sample structure. Among the surface reflectance bands, the near-infrared band 1 behaved the best performance, where the MAE was 636 kg/hm², RMSE was 784 kg/hm², and the coefficient of determination was 0.76. The near-infrared band 2 was the second most. The estimation accuracy of the blue band was the lowest, where the coefficient of determination was only 0.58, while the MAE and RMSE were 885 and 1040 kg/hm², respectively. The coefficient of determination of Normalized Difference Water Index (NDWI) was 0.82, whereas, the coefficient of determination of Normalized Difference Vegetation Index (NDVI) was 0.73. Moreover, there was a higher coefficient of determination in the combination of NDVI and NDWI, where the coefficient of determination was 0.89, and the MAE and RMSE were 444 and 527 kg/hm², respectively. There was the crucial influence of the heading stages of NDVI and NDWI on the winter wheat yield. Specifically, the accuracy of the combined NDVI and NDWI reached the highest on April 15th, indicating that the period from April 15th to 22nd presented the greatest impacts. In addition, the inter-year variation patterns from 2013 to 2015 showed well consistent trends between the estimated and observed yields in the spatial distribution, where the relative errors were within ±20%, and gradually decreased from the northwest to the southeast. A comprehensive effect of topography, temperature, and precipitation can be attributed to the surface microclimate and material redistribution. Consequently, the better yield estimation of winter wheat over the large areas can be achieved in the main winter wheat planting regions of Henan Province, using the random forest model combined with the regional frequency histogram samples. The next step can be recommended to maximize the number of samples and introduce the high-resolution satellite images for higher accuracy of the winter wheat yield estimation model. [ABSTRACT FROM AUTHOR]

Precipitation data with high resolution and high accuracy are significantly important in numerous hydrological applications. To enhance the spatial resolution and accuracy of satellite-based precipitation products, an easy-to-use downscaling-calibration method based on a spatial random forest (SRF-DC) is proposed in this study, where the spatial autocorrelation of precipitation measurements between neighboring locations is considered. SRF-DC consists of two main stages. First, the satellite-based precipitation is downscaled by the SRF with the incorporation of high-resolution variables including latitude, longitude, normalized difference vegetation index (NDVI), digital elevation model (DEM), terrain slope, aspect, relief and land surface temperatures. Then, the downscaled precipitation is calibrated by the SRF with rain gauge observations and the aforementioned high-resolution variables. The monthly Integrated MultisatellitE Retrievals for Global Precipitation Measurement (IMERG) over Sichuan Province, China, from 2015 to 2019 was processed using SRF-DC, and its results were compared with those of classical methods including geographically weighted regression (GWR), artificial neural network (ANN), random forest (RF), kriging interpolation only on gauge measurements, bilinear interpolation-based downscaling and then SRF-based calibration (Bi-SRF), and SRF-based downscaling and then geographical difference analysis (GDA)-based calibration (SRF-GDA). Comparative analyses with respect to root mean square error (RMSE), mean absolute error (MAE) and correlation coefficient (CC) demonstrate that (1) SRF-DC outperforms the classical methods as well as the original IMERG; (2) the monthly based SRF estimation is slightly more accurate than the annually based SRF fraction disaggregation method; (3) SRF-based downscaling and calibration perform better than bilinear downscaling (Bi-SRF) and GDA-based calibration (SRF-GDA); (4) kriging is more accurate than GWR and ANN, whereas its precipitation map loses detailed spatial precipitation patterns; and (5) based on the variable-importance rank of the RF, the precipitation interpolated by kriging on the rain gauge measurements is the most important variable, indicating the significance of incorporating spatial autocorrelation for precipitation estimation. [ABSTRACT FROM AUTHOR]

The total organic carbon (TOC) content is a critical parameter for estimating shale oil resources. However, common TOC prediction methods rely on empirical formulas, and their applicability varies widely from region to region. In this study, a novel data-driven Bayesian optimization extreme gradient boosting (XGBoost) model was proposed to predict the TOC content using wireline log data. The lacustrine shale in the Damintun Sag, Bohai Bay Basin, China, was used as a case study. Firstly, correlation analysis was used to analyze the relationship between the well logs and the core-measured TOC data. Based on the degree of correlation, six logging curves reflecting TOC content were selected to construct training dataset for machine learning. Then, the performance of the XGBoost model was tested using K -fold cross-validation, and the hyperparameters of the model were determined using a Bayesian optimization method to improve the search efficiency and reduce the uncertainty caused by the rule of thumb. Next, through the analysis of prediction errors, the coefficient of determination ( R 2 ) of the TOC content predicted by the XGBoost model and the core-measured TOC content reached 0.9135. The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) were 0.63, 0.77, and 12.55%, respectively. In addition, five commonly used methods, namely, Δ log R method, random forest, support vector machine, K -nearest neighbors, and multiple linear regression, were used to predict the TOC content to confirm that the XGBoost model has higher prediction accuracy and better robustness. Finally, the proposed approach was applied to predict the TOC curves of 20 exploration wells in the Damintun Sag. We obtained quantitative contour maps of the TOC content of this block for the first time. The results of this study facilitate the rapid detection of the sweet spots of the lacustrine shale oil. [ABSTRACT FROM AUTHOR]

Accurate ozone (O 3) predictions is crucial for assessing its impact on public health and developing effective prevention and control measures. While ground-based observations are considered highly accurate, the limited and uneven spatial coverage poses challenges. Air quality models can provide complete spatiotemporal coverage but are subject to biases due to simplifications in physicochemical mechanisms and uncertainties associated with emission and meteorology inputs. In this study, we aimed to improve the accuracy of the predicted O 3 concentrations from the Community Multiscale Air Quality (CMAQ) model by using the machine learning techniques. First, we compared three machine learning algorithms, namely Light Gradient Boosting Machine (LightGBM), Random Forest, and eXtreme Gradient Boosting. Results showed that LightGBM exhibited the highest correlation coefficient (R) of 0.84, making it the preferred algorithm for further analysis. Subsequently, two multi-source data prediction models based on LightGBM were constructed to improve the accuracy of predicted daily maximum 8-h O 3 (O 3 -Max8h). The first model, referred to as LGBR, utilized the predicted air pollutant concentrations and meteorological fields as input variables, while the second model, named LGBR-CHAP, incorporated ChinaHighAirPollutants (CHAP) O 3 -Max8h as an additional input variable. Validation results demonstrated significant improvements in the LGBR and LGBR-CHAP models compared to the original CMAQ model. On the daily scale, the root mean square error and mean bias of the predicted O 3 -Max8h decreased by 3.15 and 2.07 μg/m3, respectively, for the LGBR model, while they decreased by 5.61 and 4.18 μg/m3, respectively, for the LGBR-CHAP model. On the monthly scale, the R of the original CMAQ model varied from 0.2 in the March to 0.91 in June, and LGBR improved all months (0.4–0.92) as did LGBR-CHAP (0.5–0.94). Spatially, the O 3 -Max8h simulated by the original CMAQ model is better in eastern China but less skillful in western China. After optimization by the LGBR and LGBR-CHAP models, the national average R improved from 0.77 to 0.83 and 0.88, respectively. The LGBR-CHAP model exhibited better predictive capacity than the LGBR model and was subsequently employed to generate high-resolution (10 km × 10 km) and full-coverage (100%) O 3 -Max8h data. This dataset will prove valuable for future air pollution and health studies. [Display omitted] • LightGBM shows the best capacity in improving the O 3 prediction accuracy from CMAQ. • The LGBR-CHAP model has better predictive performance than the LGBR model. • The spatiotemporal pattern of O 3 -Max8h is captured (R = 0.88, RMSE = 24.4 μg/m3). • Our model generates a high-resolution (10 km × 10 km) and full-coverage (100%) O 3 data. [ABSTRACT FROM AUTHOR]

Accurate and timely maps of urban underlying land properties at the national scale are of significance in improving habitat environment and achieving sustainable development goals. Urban impervious surface (UIS) and urban green space (UGS) are two core components for characterizing urban underlying environments. However, the UIS and UGS are often mosaicked in the urban landscape with complex structures and composites. The "hard classification" or binary single type cannot be used effectively to delineate spatially explicit urban land surface property. Although six mainstream datasets on global or national urban land use and land cover products with a 30 m spatial resolution have been developed, they only provide the binary pattern or dynamic of a single urban land type, which cannot effectively delineate the quantitative components or structure of intra-urban land cover. Here we propose a new mapping strategy to acquire the multitemporal and fractional information of the essential urban land cover types at a national scale through synergizing the advantage of both big data processing and human interpretation with the aid of geoknowledge. Firstly, the vector polygons of urban boundaries in 2000, 2005, 2010, 2015 and 2018 were extracted from China's Land Use/cover Dataset (CLUD) derived from Landsat images. Secondly, the national settlement and vegetation percentages were retrieved using a sub-pixel decomposition method through a random forest algorithm using the Google Earth Engine (GEE) platform. Finally, the products of China's UIS and UGS fractions (CLUD-Urban) at a 30 m resolution were developed in 2000, 2005, 2010, 2015 and 2018. We also compared our products with six existing mainstream datasets in terms of quality and accuracy. The assessment results showed that the CLUD-Urban product has higher accuracies in urban-boundary and urban-expansion detection than other products and in addition that the accurate UIS and UGS fractions were developed in each period. The overall accuracy of urban boundaries in 2000–2018 are over 92.65 %; and the correlation coefficient (R) and root mean square errors (RMSEs) of UIS and UGS fractions are 0.91 and 0.10 (UIS) and 0.89 and 0.11 (UGS), respectively. Our result indicates that 71 % of pixels of urban land were mosaicked by the UIS and UGS within cities in 2018; a single UIS classification may highly increase the mapping uncertainty. The high spatial heterogeneity of urban underlying covers was exhibited with average fractions of 68.21 % for UIS and 22.30 % for UGS in 2018 at a national scale. The UIS and UGS increased unprecedentedly with annual rates of 1605.56 and 627.78 km2yr-1 in 2000–2018, driven by fast urbanization. The CLUD-Urban mapping can fill the knowledge gap in understanding impacts of the UIS and UGS patterns on ecosystem services and habitat environments and is valuable for detecting the hotspots of waterlogging and improving urban greening for planning and management practices. The datasets can be downloaded from 10.5281/zenodo.4034161 (Kuang et al., 2020a). [ABSTRACT FROM AUTHOR]

As a global problem, soil salinization poses a serious threat to the limited soil resources and ecosystem health in arid and semi-arid areas, and is one of the most important causes of land desertification and land degradation. Soil salinity is an effective evaluation index of soil salinization, and there is temporal and spatial difference. Dynamic monitoring can fully understand the status of soil salinization and effectively provide more quantitative information for soil restoration and land reclamation. Compared with traditional laboratory analysis, satellite remote sensing technology has major advantages in observing the ground at large spatial scales and high temporal resolution. As a new generation of spaceborne multi-spectral instrument (MSI), Sentinel-2A has novel spectral functions (namely, three red-edge bands and two near-infrared bands), which provides a broad prospect for quantitative evaluation of soil properties. At present, only a few studies were associated with edge spectrum index, vegetation index and topographic index in soil salinization mapping, and it has become a great challenge to choose the best modeling technology in soil mapping for a specific landscape area, although many learning methods have been successfully applied in the prediction of soil properties. Therefore, in this study, we used Sentinel-2A red-edge bands, vegetation indexes and digital elevation model (DEM) derived variables to conduct soil salt analysis based on machine learning methods in the Ebinur Lake wetland in the northwestern Xinjiang of China. 24 red edge spectral indices, 11 vegetation indices and 8 topographic indices were selected to participate in the modeling by the XGBoost algorithm, and the Random Forest (RF), Extreme Learning Machine (ELM) and Partial Least Squares Regressor (PLSR) three machine learning models based on 78 sampling sites were applied to extract soil Electrical Conductivity (EC). The coefficient of determination (R²), root mean square error (RMSE) and ratio of performance to deviation (RPD) were used to evaluate the prediction accuracy of the above models. The results showed that the preferred red edge spectral index could basically predict EC, and the prediction accuracy of RF prediction was the best. The verification set R², RMSE, and RPD were 0.63, 7.14 dS/m, and 2.09, respectively. The prediction accuracy of the combined modeling of the red edge spectral index and the vegetation index is better than that of the combination with the terrain index, and the prediction effect of the RF model was better than that of ELM and PLSR, and its training set (R²=0.83, RMSE=4.84 dS/m), validation set (R²=0.76, RMSE=5.36 dS/m, RPD=2.79). The prediction accuracy of the combined modeling of the red edge spectral index, vegetation index and terrain index reached the best, and RF had the best prediction effect. The R², RMSE and RPD of the verification set were 0.83, 4.81 dS/m and 3.11, respectively. In addition, with the continuous increase of input feature variables, the prediction effect of each model were improved to varying degrees. Soil salinization mapping based on the optimal variable combination (red edge spectral index + terrain index + vegetation index) and the best prediction model (RF), showed that the degree of soil salinization in the central and eastern regions was particularly serious in the study area. [ABSTRACT FROM AUTHOR]

China has experienced persistent fine particulate matter (PM 2.5) pollution for the past few years, which adversely affects both physical and mental health. The availability of high-accuracy and full-coverage PM 2.5 products will be of substantial value in formulating effective policies to combat and regulate PM 2.5 pollution. Therefore, we have developed a similarity distance-based space-time random forest (SDSTRF) model to estimate daily PM 2.5 concentrations over China by integrating surface measurements, satellite aerosol products, meteorological data, and auxiliary information. The proposed model not only accounts for spatial-temporal heterogeneity, but also uses the similarity distance to avoid errors caused by outliers. It has undergone rigorous validation through three different cross-validation (CV) approaches and has shown high and stable accuracy, particularly in the site-based CV with a coefficient of determination (R2) of 0.87, and a root mean square error (RMSE) of 10.68 μg/m3, along with a relative RMSE (rRMSE) of 27.48%. In addition, the leave-out data are predicted to determine if the SDSTRF model is accurate in its predictive power, which achieves an impressive site-based CV R2 of 0.80 and an RMSE of 12.89 μg/m3, along with an rRMSE of 33.01%. The results from a variety of validation approaches strongly indicate that the SDSTRF model can provide accurate estimation of PM 2.5 concentrations at different time scales and outperform many other space-time models by incorporating similarity distance. The proposed model would be a promising application in air pollution studies with remote sensing. • A similarity distance-based space-time random forest is used for PM 2.5 estimation. • The similarity distance indicator contributes to improving the estimation accuracy. • The model outperforms other widely used regression models in predictive power. • PM 2.5 concentrations at 0.1° are generated based on the latest MODIS AOD (C6.1). [ABSTRACT FROM AUTHOR]

• A large-area digital soil mapping method based on environmental-similarity is proposed. • The distance-decay relationship is adaptively determined by spatial distribution of samples. • Adaptively considering spatial distance to samples improves soil prediction accuracy. Digital soil mapping (DSM) based on environmental similarity may be used to predict the soil properties at unvisited locations based on the soil–environment relationship at each sample locations, which is more suitable for large-area (or regional scale) predictive mapping with comparatively limited soil samples than (geo-)statistical DSM methods with high requirements for the quantity and spatial distribution of samples. Recently, the DSM method based on environmental similarity but with inadequate consideration of the First Law of Geography was improved by considering the spatial distance to samples by means of the inverse distance weight (IDW), following the assumption that the more similar the environmental conditions and the shorter the spatial distance between the locations of interest and a sample are, the more similar the soil properties of the two locations. However, the current consideration of spatial distance by a constant distance-decay parameter should be improper across a whole large area under mapping due to the spatial variation of both sample distribution and the environmental conditions over a large area. In this paper, we propose a new large-area DSM method based on environmental similarity with adaptive consideration of spatial distance to samples by using adaptive distance-decay parameter values across a large area. An evaluation experiment to predict the soil organic matter at a depth of 0–20 cm at a 90-m resolution with a total of 659 soil samples in Anhui Province (approximately 134000 km2), China, was carried out. Evaluation results show that the proposed method (named iPSM-AdaIDW) achieved higher accuracy (with a root mean square error, or RMSE, of 8.273 g/kg) than two representative environmental-similarity-based DSM methods, i.e., individual Predictive Soil Mapping (iPSM) (without considering the spatial distance to samples) (RMSE = 8.878 g/kg) and iPSM considering spatial distance with a constant distance-decay parameter across the study area (RMSE = 8.310 g/kg), and a representative of large-area DSM not based on environmental similarity, i.e., the random forest kriging method (RMSE = 9.647 g/kg). For large areas with quantitively limited and unevenly distributed samples, the proposed method can achieve more applicable and accurate predictions. [ABSTRACT FROM AUTHOR]

According to frontier settlement theory, physical topography is a critical factor that shapes a region's psychological landscape. To determine whether physical topography also impacts generalized trust, we conducted an exploratory study to investigate the relationship between mountainousness and individual's generalized trust (N = 26,910) in China. The multilevel modeling analysis revealed positive associations between the standard deviation of elevation and generalized trust. Random Forest analyses confirmed mountainousness indicators as significant predictors of generalized trust and also uncovered a complex nonlinear relationship between these indicators and generalized trust. Our results revealed that individuals residing in China's mountainous areas are more prone to trust others. • The present research investigates the association between mountainousness and generalized trust in China. • The standard deviation of elevation is positively correlated with generalized trust. • Random forest analyzes reveal nonlinear relationships between mountainousness indicators and generalized trust. • Individuals residing in China's mountainous areas are more prone to trust others. [ABSTRACT FROM AUTHOR]

High-durability concrete is required in extremely cold or ocean environments, making the design of concrete mixes highly important and complicated. In this study, a hybrid intelligent framework for multi-objective optimization based on random forest (RF) and the non-dominated sorting genetic algorithm version II (NSGA-II) is developed to efficiently predict concrete durability and optimize the concrete mix ratio. The relative dynamic elastic modulus of concrete after 300 freeze–thaw cycles and the chloride ion permeability coefficient at 28 days are defined as the standard measures of durability. The concrete mix ratio is taken as the influencing parameter, and orthogonal test data and engineering practice data are collected as the datasets. The proposed framework is applied to a realistic expressway project in a cold region of China. The results demonstrate that (1) a hybrid intelligent framework based on RF-NSGA-II can effectively predict concrete durability and optimize the mix ratio. (2) The developed RF model has an excellent regression learning ability, while the goodness of fit (R2) of concrete durability reaches 0.9503 and 0.9551, respectively, with root mean square error (RMSE) values of only 0.096 and 0.043, the mean absolute percentage error (MAPE) values of 2.54% and 2.17%. (3) After optimization, the concrete durability reaches a high standard, with a frost resistance of >95% and a chloride ion permeability coefficient of <3*10 − 8 cm2/s, at a unit volume cost of only 376.77 yuan. Hence, the proposed framework can be used to effectively optimize the concrete mix design and provide guidance for similar projects. [ABSTRACT FROM AUTHOR]

• The Random Forest model is applied to predict the CO 2 emission caused by road transport. • Six Meta-Heuristic Algorithms are proposed to optimize the hyperparameters of the Random Forest model. • The hybrid RF-SMA model has the best performance in improving predictions. • The R2 metric for the proposed model is 0.9942 for train data and 0.9641 for test data. Carbon dioxide is one of the most important pollutants in urban areas. Since the relationship between the factors of road transport and CO 2 emission is often complex, using methods based on computational intelligence can be useful. In this way, a hybrid Random Forest, support vector regression, and response surface methodology are implemented to predict CO 2 emission in 30 major cities in China. Also, seven optimizers are applied to the Random Forest, and two optimizers are applied to the support vector regression methods to tune their hyper-parameters. The mentioned methods' accuracy is compared through the Standard Error (SE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Relative Absolute Error (RAE), and the coefficient of determination (R2) statistical indexes. The obtained results reveal that the support vector regression with Harris Hawk optimizer has the best accuracy in the training process with an R2 value of 0.9999, and the Random Forest with the Slime Mould Algorithm with an R2 value of 0.9641 has the best accuracy in the testing process. Hence, the Random Forest with Slime Mould Algorithm (RF-SMA) is the best method to predict CO 2 emission. [ABSTRACT FROM AUTHOR]

[Display omitted] • Three hybrid RF models were developed to predict blast-induced overbreak. • The RF-TSA model outperformed other predictive RF models in estimating overbreak. • The blast design parameters showed a deep impact on the overbreak generation. • The PF was selected as the most influential factor in the overbreak generation. Overbreak is a detrimental phenomenon caused by tunnel blasting, which can lead to increased time and cost in the construction schedule. It is very important to establish a model that can accurately predict the overbreak caused by tunnel blasting. To achieve this goal, the random forest (RF) is an ensemble machine learning model optimised by three metaheuristic algorithms to predicted overbreak, i.e. the grey wolf optimiser (GWO), the whale optimisation algorithm (WOA), and the tunicate swarm algorithm (TSA). The primary roles of GWO, WOA, and TSA are to search for the optimal hyper-parameters of the RF model in the solution space. To create the models above, 523 data samples were taken from a highway tunnel in China. The established database comprised seven predictors or inputs, including the number of holes, hole depth, total charge, advance length, rock mass rating, tunnel cross-sectional area, and powder factor. Three hybrid RF-based models (RF-GWO, RF-WOA, and RF-TSA) were constructed to predict overbreak. Subsequently, the performance levels of the developed hybrid models were evaluated according to four indices: the coefficient of determination, the root mean square error, the variance accounted for, and the A-20 index. The results showed that the TSA optimisation algorithm was better than the other two algorithms (WOA and GWO) at finding the best hyper-parameters for the RF model. Moreover, the results of comparative analysis with the single RF model confirmed that the proposed RF-TSA model is a strong solution with high accuracy for tackling the overbreak issues. The results of this study showed that the developed models can provide more accurate overbreak values compared to the intelligent techniques available in the literature; they can be used in practice and similar projects. [ABSTRACT FROM AUTHOR]

• Pretreatment combined algorithms combining multiple mathematical transformations and dimensionality reduction algorithms to comprehensively process raw spectral data. • The advantage of this method is high accuracy". • A method for selecting the optimal Pretreatment combined algorithm is proposed and can accurately predict the accuracy of skarn-type iron ore. Iron ore grade determination is a key step in mining production, and its efficiency has a significant impact on mining efficiency and profitability. At present, the most commonly used method for ore grade determination is chemical analysis, which not only is expensive but also has a long assay cycle and lag effect relative to the ore allocation process and cannot effectively reduce the loss depletion rate of ore mining. Therefore, in situ iron ore grade data determination technology based on visible and near-infrared hyperspectral data analysis is an effective way to solve the above problems, but the poor quality of raw hyperspectral data and low accuracy of the ore grade inversion model are still one of the bottlenecks in the application of this technology. In this paper, 225 ore samples were collected from Hongling skarn-type iron ore in Chifeng, Inner Mongolia Autonomous Region, China. The visible and near-infrared hyperspectral data of the sample set and the grade data obtained from chemical analysis were used as data sources. The raw spectral data were first smoothed, and then the smoothed spectral data were processed using two algorithms, continuum removal (CR) and standard normal variate transform (SNV). Then, two dimensionality reduction algorithms, principal component analysis (PCA) and genetic algorithm (GA), were used to reduce the dimensionality of the spectral data before and after preprocessing, and the data sources were obtained after the four different preprocessing combination algorithms were processed. Finally, several quantitative inversion models were established based on the random forest algorithm and the backpropagation neural network (BPNN) algorithm. The stability, accuracy and credibility of the models were evaluated by four metrics: the coefficient of determination (R2), root mean square error (RMSE), residual prediction deviation(RPD) and mean relative error (MRE). The results showed that the quantitative inversion model (SNV-PCA-BPNN) established based on the BP neural network algorithm using the data processed by SNV and PCA for dimensionality reduction was the best, with an R2 of 0.99, an RMSE of 0.0048, an MRE of 1.10% and a RPD of 37.70, providing an effective method for the high-precision inversion of iron ore grades. [ABSTRACT FROM AUTHOR]

• Monitored the spatial heterogeneity and driving mechanism of aboveground biomass (AGB) by combining Space-Air-Ground integrated and Random Forest Regression (RFR) model. • Pointed out the main indicators for estimating AGB in different coverage regions. • Assessed the applicability of RFR model for estimating the AGB in three different coverage regions. Accurate monitoring of aboveground biomass (AGB) and its spatial distribution is essential for sustainable management and development. The most frequently used method for estimating AGB with remotely sensed data is based on the relationship between AGB measured data and vegetation indices. In the past, there were some shortcomings in analyzing the spatial heterogeneity and solving scale difference. Therefore, this paper realized the space-air-ground integrated based on Sentinel-2 Multispectral Instrument (MSI), Unmanned Aerial Vehicle (UAV) and measured data. The results indicated that the RFR models have excellent adaptability to invert AGB of three different coverage regions in study area, with R2 of 0.69 and root mean square error (RMSE) of 116.22 kg·hm−2 for low-coverage region, R2 of 0.65 and RMSE of 164.16 kg·hm−2 for medium-coverage region, R2 of 0.73 and RMSE of 149.99 kg·hm−2 for high-coverage region. We found that the main driving vegetation indices was different in each region during the process of model building and it was related to the characteristics of vegetation indices. Moreover, regions with different coverage have different responses to topography. The findings showed that the spatial distribution and driving mechanism of AGB in the study area could be more clearly analyzed based on the coverage division. The introduction of coverage division reasonably enhanced the accuracy of RFR modeling AGB. This study confirmed that using the Sentinel-2 MSI, UAV images, measured data and the RFR model to realize the space-air-ground integrated is an applicable method for monitoring and assessment of AGB in Qilian Mountains, which can offer scientific basis and theoretical support to rational development of land resources and maintaining the balance of ecosystem. [ABSTRACT FROM AUTHOR]

• Filter, wrapper, embedded and ensemble feature selection approaches (FS) were evaluated. • The proposed ensemble FS incorporates ten individual selectors to identify top significant variables. • Ensemble FS produced the highest predictive variable subset for mapping SOM. • XGBoost outperformed RF in mapping SOM in subtropical restored forests. Mapping Soil organic matter (SOM) over a complex forest landscape is challenging due to the difficulty in selecting the most insightful variables from high-dimensional datasets in the recent explosion of geospatial-data. Feature selection (FS) is necessary to reduce data redundancy and noise as well as to achieve more reliable SOM spatial predictions. However, it is still unclear that which is most effective among various FS methods in mapping SOM. Therefore, four types of FS approaches (i.e., filter, wrapper, embedded and ensemble) were adopted to generate optimum variable subsets from an original variable dataset of 60 candidates, respectively, for mapping SOM of restored forest land in a typical subtropical region of southern China. The most used methods for each type of FS approaches were selected in this study, including three filters (Chi-square, InfoGain and pearson correlation analysis), three wrappers (genetic algorithm, simulated annealing algorithm and support vector machine-recursive feature elimination) and three embedded methods (Boruta, random forest (RF) and extreme gradient boosting (XGBoost)), as well as an ensemble method (robust rank aggreg algorithm (RRA)). Meanwhile, the RF and XGBoost models were applied with a 10-fold cross-validation method to compare the relative advantages of the different FS methods in SOM mapping, by utilizing the correlation coefficients R2 between observed and predicted values and predicting errors of root mean square error (RMSE). The results show that the SOM prediction accuracies with optimized variable subsets generated by the different FS methods are better than those with full variables, yet the improvements of prediction performance are different among the four types of FS approaches. The ensemble method (RRA) is superior to the other three types of approaches with an average RMSE reduction of 9.16% comparing to that without using FS methods, followed by wrapper and embedded methods which obtained the average RMSE reduction by 7.81%, 7.32%, respectively, and the filter methods are the weakest in the RMSE reduction with slight decreases of 4.32%. The XGBoost model achieved a better performance in predicting SOM than the RF model regardless of input variables, and the XGBoost model combined with RRA FS method shows the greatest potential to map SOM in the restored forest land. This study provides a reference for obtaining more parsimonious and robust variable sets from the available big geo-data freely for soil mapping in other areas. [ABSTRACT FROM AUTHOR]

Satellite retrieval can offer global soil moisture information, such as Advanced Microwave Scanning Radiometer-Earth Observing System (AMSR-E) data. AMSR-E has been used to provide soil moisture all over the world, with a coarse resolution of 25 km × 25 km. The coarse resolution of the soil moisture dataset often hinders its use in local or regional research. This work proposes a new framework based on the random forest (RF) model while using five auxiliary data to downscale the AMSR-E soil moisture data over North China. The downscaled results with a 1 km spatial resolution are verified against in situ measurements. Compared with AMSR-E data, the correlation coefficient of the downscaled data is increased by 0.17, and the root mean squared error, mean absolute error, and unbiased root mean square error are reduced by 0.02, 0.01, and 0.03 m3/m3, respectively. In addition, the comparison results with Multiple Linear Regression and Support Vector Regression downscaled data show that the proposed method significantly outperforms the other two methods. The feasibility of our model is well supported by the importance analysis and leave-one-out analysis. Our study, which combines RF with spatiotemporal search algorithms and efficient auxiliary data, may provide insights into soil moisture downscaling in large areas with various surface characteristics and climatic conditions. [ABSTRACT FROM AUTHOR]

Burn severity is a key component of fire regimes and is critical for quantifying fires' impacts on key ecological processes. The spatial and temporal distribution characteristics of forest burn severity are closely related to its environmental drivers prior to the fire occurrence. The temperate coniferous forest of northern China is an important part of China's forest resources and has suffered frequent forest fires in recent years. However, the understanding of environmental drivers controlling burn severity in this fire-prone region is still limited. To fill the gap, spatial pattern metrics including pre-fire fuel variables (tree canopy cover (TCC), normalized difference vegetation index (NDVI), and live fuel moisture content (LFMC)), topographic variables (elevation, slope, and topographic radiation aspect index (TRASP)), and weather variables (relative humidity, maximum air temperature, cumulative precipitation, and maximum wind speed) were correlated with a remote sensing-derived burn severity index, the composite burn index (CBI). A random forest (RF) machine learning algorithm was applied to reveal the relative importance of the environmental drivers mentioned above to burn severity for a fire. The model achieved CBI prediction accuracy with a correlation coefficient (R) equal to 0.76, root mean square error (RMSE) equal to 0.16, and fitting line slope equal to 0.64. The results showed that burn severity was mostly influenced by flammable live fuels and LFMC. The elevation was the most important topographic driver, and meteorological variables had no obvious effect on burn severity. Our findings suggest that in addition to conducting strategic fuel reduction management activities, planning the landscapes with fire-resistant plants with higher LFMC when possible (e.g., "Green firebreaks") is also indispensable for lowering the burn severity caused by wildfires in the temperate coniferous forests of northern China. [ABSTRACT FROM AUTHOR]

[Display omitted] • Proposed a framework for detecting and quantifying surface SOC content in complex terrain. • The effects of different ancillary variables in modelling were determined. • Recursive feature elimination ensures the integrity and robustness of the inputs. • Texture features were efficient and flexible for enriching the RF regression model. • Regional soil organic carbon was mapped with GF-5 satellite hyperspectral data. The prediction of soil organic carbon (SOC) from hyperspectral data often lacks geographic and environmental information related to soil genesis, which would improve the accuracy of the predicted SOC. The main purpose of this study was to improve the accuracy of SOC prediction and the mapping of SOC spatial distributions. We employed satellite hyperspectral image (HSI) data combined with ancillary variables (spectral indexes (SIs), terrain attributes (TAs) and spectral texture features (TFs)) by first stratifying the soil at the great group level. The central part of the Songnen Plain in Northeast China was selected as a region for a case study, because the region attracts considerable research interest as major grain production area in China. In different prediction models, recursive feature elimination (RFE) was applied to optimize input variables to reflect the soil-landscape relationships of different soil classes. The results showed that when the soil stratification strategy and ancillary variables were comprehensively considered, the accuracy of the model was significantly improved (with a coefficient of determination (R2) of 0.76, root mean square error (RMSE) of 3.16 g kg−1, and ratio of performance to interquartile distance (RPIQ) of 2.28). The introduction of SIs, TAs and TFs improved the R2 values by 6.15%, 6.15%, and 13.85%, respectively, compared to those achieved with the original reflectance (OR) bands alone. Moreover, the introduction of ancillary variables improved the accuracies of the SOC models, yielding R2 values of Phaeozems, Chernozems, Arenosols and Cambisols of 0.79, 0.53, 0.76, and 0.81, respectively. Compared with the prediction model, which is based on only the OR, the proposed model can better explain SOC spatial variations. The performance comparison highlights the advantage of the considering geomorphic features when utilized for SOC prediction in regional-scale; this model covers the elimination and expression of optimal ancillary variables for different soil classes, which are closely related to the formation of various soil types and the geomorphic evolution of the region. The SOC map that we obtained shows detailed soil information and effectively expresses the soil factors associated with the environment. The map can support planners in establishing efficient SOC monitoring methods and assessments and prioritizing inputs for future exploitation and research. [ABSTRACT FROM AUTHOR]

Spectral unmixing remains the most popular method for estimating the composition of mixed pixels. However, the spectral-based unmixing method cannot easily distinguish vegetation with similar spectral characteristics (e.g., different forest tree species). Furthermore, in large areas with significant heterogeneity, extracting a large number of pure endmember samples is challenging. Here, we implement a fractional evergreen forest cover-self-adaptive parameter (FEVC-SAP) approach to measure FEVC at the regional scale from continuous intra-year time-series normalized difference vegetation index (NDVI) values derived from moderate resolution imaging spectroradiometer (MODIS) imagery acquired over southern China, an area with a complex mixture of temperate, subtropical, and tropical climates containing evergreen and deciduous forests. Considering the cover of evergreen forest as a fraction of total forest (evergreen forest plus non-evergreen forest), the dimidiate pixel model combined with an index of evergreen forest phenological characteristics (NDVIann-min: intra-annual minimum NDVI value) was used to distinguish between evergreen and non-evergreen forests within a pixel. Due to spatial heterogeneity, the optimal model parameters differ among regions. By dividing the study area into grids, our method converts image spectral information into gray level information and uses the Otsu threshold segmentation method to simulate the appropriate parameters for each grid for adaptive acquisition of FEVC parameters. Mapping accuracy was assessed at the pixel and sub-pixel scales. At the pixel scale, a confusion matrix was constructed with higher overall accuracy (87.5%) of evergreen forest classification than existing land cover products, including GLC 30 and MOD12. At the sub-pixel scale, a strong linear correlation was found between the cover fraction predicted by our method and the reference cover fraction obtained from GF-1 images (R2 = 0.86). Compared to other methods, the FEVC-SAP had a lower estimation deviation (root mean square error = 8.6%). Moreover, the proposed method had greater estimation accuracy in densely than sparsely forested areas. Our results highlight the utility of the adaptive-parameter linear unmixing model for quantitative evaluation of the coverage of evergreen forest and other vegetation types at large scales. [ABSTRACT FROM AUTHOR]

• Develop a within-growing season yield forecast system with random forest model. • Random forest model performs well in predicting grain yield in China. • We identified the most important stage-specific predictors determining crop yield. • The most important variable influencing yields varied with crop types. Accurate and timely crop yield forecasts can provide essential information to make conclusive agricultural policies and to conduct investments. Recent studies have used different machine learning techniques to develop such yield forecast systems for single crops at regional scales. However, no study has used multiple sources of environmental predictors (climate, soil, and vegetation) to forecast yields for three major crops in China. In this study, we adopted 7-year observed crop yield data (2013–2019) for three major grain crops (wheat, maize, and rice) across China, and three major data sets including climate, vegetation indices, and soil properties were used to develop a dynamic yield forecasting system based on the random forest (RF) model. The RF model showed good performance for estimating yields of all three crops with correlation coefficient (r) higher than 0.75 and normalized root means square errors (nRMSE) lower than 18.0%. Our results also showed that crop yields can be satisfactorily forecasted at one to three months prior to harvest. The optimum lead time for yield forecasting depended on crop types. In addition, we found the major predictors influencing crop yield varied between crops. In general, solar radiation and vegetation indices (especially during jointing to milk development stages) were identified as the main predictor for winter wheat; vegetation indices (throughout the growing season) and drought (especially during emergence to tasseling stages) were the most important predictors for spring maize; soil moisture (throughout the growing season) was the dominant predictor for summer maize, late rice, and mid rice; precipitation (especially during booting to heading stages) was the main predictor for early rice. Our study provides insights into practical crop yield forecasting and the understanding of yield response to environmental conditions at a large scale across China. The methods undertaken in this research can be easily implemented in other countries with available information on climate, soil, and vegetation conditions. [ABSTRACT FROM AUTHOR]

Passive microwave surface soil moisture (SSM) products tend to have very low resolution, which massively limits their application and validation in regional or local-scale areas. Many climate and hydrological studies are urgently needed to evaluate the suitability of satellite SSM products, especially in alpine mountain areas where soil moisture plays a key role in terrestrial atmospheric exchanges. Aiming to overcome this limitation, a downscaling method based on random forest (RF) was proposed to disaggregate satellite SSM products. We compared the ability of the downscaled soil moisture active passive (SMAP) SSM and soil moisture and ocean salinity satellite (SMOS) SSM products to capture soil moisture information in upstream of the Heihe River Basin by using in situ measurements, the triple collocation (TC) method and temperature vegetation dryness index (TVDI). The results showed that the RF downscaling method has strong applicability in the study area, and the downscaled results of the two products after residual correction have more details, which can better represent the spatial distribution of soil moisture. The validation with the in situ SSM measurements indicates that the correlation between downscaled SMAP and in situ SSM is better than downscaled SMOS at both point and watershed scales in the Babaohe River Basin. From the TC method, the root mean square error (RMSE) of the CLDAS (CMA land data assimilation system), downscaled SMAP and downscaled SMOS were 0.0265, 0.0255 and 0.0317, respectively, indicating that the downscaled SMAP has smaller errors in the study area than others. However, the soil moisture distribution in the study area shown by the SMOS downscaled results is closer than the downscaled SMAP to the degree of drought reflected by TVDI. Overall, this study suggests that the proposed RF-based downscaling method can capture the variation of SSM well, and the downscaled SMAP products perform significantly better than the downscaled SMOS products after the accuracy verification and error analysis of the downscaled results, and it should be helpful to facilitate applications for satellite SSM products at small scales. [ABSTRACT FROM AUTHOR]

Digital soil mapping approaches related to soil organic matter (SOM) are crucial to quantify the process of the carbon cycle in terrestrial ecosystems and thus, can better manage soil fertility. Recently, many studies have compared machine learning (ML) models with traditional statistical models in digital soil mapping. However, few studies focused on the application of hybrid models that combine ML with statistical models to map SOM content, especially in loess areas, which have a complicated geomorphologic landscape. In this study, the trend prediction used two ML models, i.e., gradient boosting modeling and random forest (RF), and a traditional stepwise multiple linear regression plus interpolated residuals generated from two classic geostatistical models, i.e., ordinary kriging and inverse distance weighting, and a high accuracy surface modeling (HASM) were implemented to map SOM content in the Dongzhi Loess Tableland area of China. A total of 145 topsoil samples and heterogeneous environmental variables were collected to develop the hybrid models. Results showed that 18 variables related to soil properties, climate variables, terrain attributes, vegetation indices, and location attributes played an important role in SOM mapping. The models that incorporate ML algorithms and interpolated residuals to predict SOM variation were found to have a better ability to handle complex environment relationships. The HASM model outperformed traditional geostatistical models in interpolating the residuals. In contrast, RF combined with HASM residuals (RF_HASM) gave the best performance, with the lowest mean absolute error (1.69 g/kg), root mean square error (2.30 g/kg), and the highest coefficient of determination (0.57) and concordance correlation coefficient (0.69) values. Moreover, the spatial distribution pattern obtained with RF_HASM yielded a spatial distribution of SOM that better fit the actual distribution pattern of the study area. In conclusion, these results suggest that RF_HASM is particularly capable of improving the mapping accuracy of SOM content at the regional scale. [ABSTRACT FROM AUTHOR]

Accurate prediction of dissolved oxygen time series is important for improving the water environment and aiding water resource management. In this study, four stand-alone models including multiple linear regression (MLR), support vector machine (SVM), artificial neural network (ANN) and random forest (RF), and four hybrid models based on wavelet transform (WT) including WT-MLR, WT-SVM, WT-ANN and WT-RF were used to predict the daily dissolved oxygen (DO) at 1–5-day lead times in the Dongjiang River Basin, China. To make the prediction robust, the maximal information coefficient (MIC) was used to capture comprehensive information between DO and explanatory variables. The 5-fold cross validation grid search approach was used to optimize parameters of machine learning tools. Two types of frameworks of WT: direct framework (i.e., only the explanatory variables were decomposed) and multicomponent framework (i.e., both explanatory variables and target variables were decomposed) were used to construct hybrid models. The results show that MIC extracts four optimal explanatory variables: previous DO, water temperature, air temperature and air pressure. Four evaluation parameters including correlation coefficient (R), Nash-Sutcliffe efficiency (NSE), mean absolute error (MAE) and root mean square error (RMSE) indicate that the prediction accuracy decreases as the lead time changes from 1 to 5 days. In terms of the stand-alone models, MLR model outperforms the other three models with higher NSE values of 0.616–0.921, and lower RMSE values of 0.503–1.111. With regard to the hybrid models, WT-ANN and WT-MLR models exhibit higher performance, and multicomponent framework performs better than direct framework in all hybrid models. In general, the multicomponent framework of WT can improve the prediction accuracy of stand-alone models at a certain degree, while the direct framework shows no obvious advantage. • Using stand-alone models and hybrid wavelet-based models to estimate river dissolved oxygen levels at 1–5-day lead times. • Extracting four explanatory variables using maximal information coefficient. • Hybrid wavelet-based models outperform stand-alone models at a certain degree. • The multicomponent framework of wavelet transform outperforms the direct framework. [ABSTRACT FROM AUTHOR]

In order to explore the spatiotemporal changes and driving factors of soil organic carbon (SOC) in the agro-pastoral ecotone of northern China, we took Aohan banner, Chifeng City, Inner Mongolia Autonomous Region as the study area, used the random forest (RF) method to predict the SOC from 1989 to 2018, and the geographic detector method (GDM) was applied to analyze quantitatively the natural and anthropogenic factors that are affecting Aohan banner. The results indicated that: (1) After adding the terrain factors, the R2 and residual predictive deviation (RPD) of the RF model increased by 1.178 and 0.39%, with root mean square errors (RMSEs) of 1.42 g/kg and 1.05 g/kg, respectively; (2) The spatial distribution of SOC was higher in the south and lower in the north; the negative growth of SOC accounted for 55.923% of the total area, showing a trend of degradation; (3) Precipitation was the main driving factor of SOC spatial variation in the typical agro-pastoral ecotone of northern China, which was also affected by temperature, elevation, soil type and soil texture (p < 0.01). (4). Anthropogenic factors (carbon input and gross domestic product (GDP)) had a greater impact on SOC than did climate factors (temperature and precipitation), making anthropogenic factors the dominant factors affecting SOC temporal variation (p < 0.01). The results of this work constitute a basis for a regional assessment of the temporal evolution of organic carbon in the soil surface, which is a key tool for monitoring the sustainable development of agropastoral ecotones. [ABSTRACT FROM AUTHOR]

Accurate measurement of forest growing stem volume (GSV) is important for forest resource management and ecosystem dynamics monitoring. Optical remote sensing imagery has great application prospects in forest GSV estimation on regional and global scales as it is easily accessible, has a wide coverage, and mature technology. However, their application is limited by cloud coverage, data stripes, atmospheric effects, and satellite sensor errors. Combining multi-sensor data can reduce such limitations as it increases the data availability, but also causes the multi-dimensional problem that increases the difficulty of feature selection. In this study, GaoFen-2 (GF-2) and Sentinel-2 images were integrated, and feature variables and data scenarios were derived by a proposed adaptive feature variable combination optimization (AFCO) program for estimating the GSV of coniferous plantations. The AFCO algorithm was compared to four traditional feature variable selection methods, namely, random forest (RF), stepwise random forest (SRF), fast iterative feature selection method for k-nearest neighbors (KNN-FIFS), and the feature variable screening and combination optimization procedure based on the distance correlation coefficient and k-nearest neighbors (DC-FSCK). The comparison indicated that the AFCO program not only considered the combination effect of feature variables, but also optimized the selection of the first feature variable, error threshold, and selection of the estimation model. Furthermore, we selected feature variables from three datasets (GF-2, Sentinel-2, and the integrated data) following the AFCO and four other feature selection methods and used the k-nearest neighbors (KNN) and random forest regression (RFR) to estimate the GSV of coniferous plantations in northern China. The results indicated that the integrated data improved the GSV estimation accuracy of coniferous plantations, with relative root mean square errors (RMSErs) of 15.0% and 19.6%, which were lower than those of GF-2 and Sentinel-2 data, respectively. In particular, the texture feature variables derived from GF-2 red band image have a significant impact on GSV estimation performance of the integrated dataset. For most data scenarios, the AFCO algorithm gained more accurate GSV estimates, as the RMSErs were 30.0%, 23.7%, 17.7%, and 17.5% lower than those of RF, SRF, KNN-FIFS, and DC-FSCK, respectively. The GSV distribution map obtained by the AFCO method and RFR model matched the field observations well. This study provides some insight into the application of optical images, optimization of the feature variable combination, and modeling algorithm selection for estimating the GSV of coniferous plantations. [ABSTRACT FROM AUTHOR]

• A method of extracting tree height using DSM of stereo image and DEM is proposed. • Build different AGB estimation models through incorporation of tree height variable. • The forests AGB estimation model with tree height alleviates the saturation problem. The forest tree height and aboveground biomass (AGB) are important indicators for monitoring changes and trends in forest carbon storage and terrestrial carbon fluxes. Accurate large-scale wall-to-wall mapping of the forest tree height and AGB remain challenging due to the limited data availability for extraction tree height and the data signal saturation problem in AGB estimation. In this study, we explored the potential of forest tree height mapping using stereo imageries, and analyzed whether accounting for such information, in addition to optical sensor data, could improve the performance of AGB estimations of coniferous forests in a case study in North China. First, a spatially continuous tree height product was obtained using Ziyuan-3 satellite (ZY-3) stereo images combined with a digital elevation model (DEM) obtained from Advanced Land Observing Satellite (ALOS) data. Second, two AGB estimation models were established by combining the forest tree height with vegetation index, spectral, biophysical (from Sentinel-2 images), and topographic variables. A random forest algorithm was utilized to evaluate the effect of including the tree height variable in the AGB estimation. The results showed that the tree height estimation using the nadir and forward views of the ZY-3 stereo images was more accurate than that based on the nadir and backward views from the same images. The AGB estimation model incorporating the tree height variable with a coefficient of determination value of 0.7789, a root mean square error (RMSE) value of 29.815 Mg/ha and a relative RMSE of 23.42% was more robust and effective, thereby demonstrating that the tree height variable can be used to alleviate the data signal saturation issue successfully. The proposed approach can provide new insight into forest tree height mapping and AGB products obtained from satellite stereo images and freely accessible Sentinel-2 multispectral images. [ABSTRACT FROM AUTHOR]

Software Project Estimation is a challenging and important activity in developing software projects. Software Project Estimation includes Software Time Estimation, Software Resource Estimation, Software Cost Estimation, and Software Effort Estimation. Software Effort Estimation focuses on predicting the number of hours of work (effort in terms of person-hours or person-months) required to develop or maintain a software application. It is difficult to forecast effort during the initial stages of software development. Various machine learning and deep learning models have been developed to predict the effort estimation. In this paper, single model approaches and ensemble approaches were considered for estimation. Ensemble techniques are the combination of several single models. Ensemble techniques considered for estimation were averaging, weighted averaging, bagging, boosting, and stacking. Various stacking models considered and evaluated were stacking using a generalized linear model, stacking using decision tree, stacking using a support vector machine, and stacking using random forest. Datasets considered for estimation were Albrecht, China, Desharnais, Kemerer, Kitchenham, Maxwell, and Cocomo81. Evaluation measures used were mean absolute error, root mean squared error, and R-squared. The results proved that the proposed stacking using random forest provides the best results compared with single model approaches using the machine or deep learning algorithms and other ensemble techniques. [ABSTRACT FROM AUTHOR]

In this paper, based on electrical power consumption (EPC) data extracted from DMSP/OLS night light data, we select three national-level urban agglomerations in China's Yangtze River Economic Belt(YREB), includes Yangtze River Delta urban agglomerations(YRDUA), urban agglomeration in the middle reaches of the Yangtze River(UAMRYR), and Chengdu-Chongqing urban agglomeration(CCUA) as the research objects. In addition, the coefficient of variation (CV), kernel density analysis, cold hot spot analysis, trend analysis, standard deviation ellipse and Moran's I Index were used to analyze the Spatio-temporal Dynamic Evolution Characteristics of EPC in the three urban agglomerations of the YREB. In addition, we also use geographically weighted regression (GWR) model and random forest algorithm to analyze the influencing factors of EPC in the three major urban agglomerations in YREB. The results of this study show that from 1992 to 2013, the CV of the EPC in the three urban agglomerations of YREB has been declining at the overall level. At the same time, the highest EPC value is in YRDUA, followed by UAMRYR and CCUA. In addition, with the increase of time, the high-value areas of EPC hot spots are basically distributed in YRDUA. The standard deviation ellipses of the EPC of the three urban agglomerations of YREB clearly show the characteristics of "east-west" spatial distribution. With the increase of time, the correlations and the agglomeration of the EPC in the three urban agglomerations of the YREB were both become more and more obvious. In terms of influencing factor analysis, by using GWR model, we found that the five influencing factors we selected basically have a positive impact on the EPC of the YREB. By using the Random forest algorithm, we found that the three main influencing factors of EPC in the three major urban agglomerations in the YREB are the proportion of secondary industry in GDP, Per capita disposable income of urban residents, and Urbanization rate. [ABSTRACT FROM AUTHOR]

Much of the population is exposed to PM2.5 (particulate matter) pollution in China, and establishing a high-precision PM2.5 grid dataset will be very valuable for air pollution and related studies. However, limited by the traditional models themselves and input data sources, PM2.5 estimations are of low accuracy with narrow spatial coverage. Therefore, we develop a new spatiotemporally weighted random forest (SWRF) model to improve the estimation accuracy and expand the spatial coverage of PM2.5 concentrations using the latest release of the Visible infrared Imaging Radiometer (VIIRS) Deep Blue (DB) aerosol product, along with meteorological variables, and socioeconomic data. Compared with traditional methods and the results of previous similar studies, our satellite-derived PM2.5 distribution shows better consistency with surface-measured records, having a high out-of-sample (out-of-station) cross-validation (CV) coefficient of determination (CV-R2), root mean squared error (RMSE), and mean absolute error (MAE) of 0.87 (0.85), 11.23 (11.53) μg m−3 and 8.25 (8.78) μg m−3, respectively. The monthly, seasonal, and annual mean PM2.5 were also successfully captured (CV-R2 = 0.91–0.92, RMSE = 4.35–6.72 μg m−3). Then, the spatial characteristics of PM2.5 pollution in 2018 were investigated, showing that although air pollution has diminished in recent years, China still faces a high PM2.5 pollution risk overall, especially in winter (average = 50.43 + 16.81 μg m−3). In addition, 19 provinces or administrative regions have annual PM2.5 concentrations >35 μg m−3, particularly the Xinjiang Uygur Autonomous Region (~55.25 μg m−3), Tianjin (~49.65 μg m−3), and Henan Province (~48.60 μg m−3). Our estimated surface PM2.5 concentrations are accurate, which could benefit further research on air pollution in China. [ABSTRACT FROM AUTHOR]

Fine particulate matter such as PM 2.5 has been the focus of increasing public concerns because of its adverse effect on environment and health risks. However, existing efforts of mapping PM 2.5 concentrations are always limited by coarse spatial resolutions and temporal frequencies. Addressing this shortcoming, here we explicitly estimated hourly PM 2.5 concentrations at 1-km spatial resolution in China from March 2018 to February 2019 using a two-stage random forest model. In the first stage, we used a gap-filling method to generate full-coverage Aerosol Optical Depth (AOD) by fusing AOD data from satellite (Himawari-8 and MODIS) and weather forecast model (CAMS), and additional meteorological and geographical variables. Gap-filled AOD was subsequently used to estimate hourly PM 2.5 in the Stage II. Results showed that our model achieved accurate and robust estimations of PM 2.5 concentrations, with an overall cross-validated R2 of 0.85, root mean squared error of 11.02 μg/m3, and mean absolute error of 6.73 μg/m3. CAMS-simulated PM 2.5 , elevation, and gap-filled AOD were identified to be important variables contributing to the model performance of PM 2.5 estimation. The model performance varied over the daily temporal scale. Specifically, daily estimation model performed better in spring and winter but worse in summer and autumn. We provide an alternative to generate spatially and temporally explicit mapping of PM 2.5 concentrations with fine resolutions, making it possible to achieve real-time monitoring of air pollutions. The detailed spatial heterogeneity and diurnal variability of PM 2.5 concentrations will also be valuable for environmental exposure assessments. Unlabelled Image • A two-stage random forest model was developed to estimate hourly full-coverage PM2.5 at 1 km resolution in China. • Gap-filled AOD was generated using AHI, MAIAC and CAMS AOD. • Estimation of PM2.5 concentrations achieved relatively high accuracies. • CAMS PM2.5 simulations, elevation and gap-filled AOD greatly contributed to the model performance. [ABSTRACT FROM AUTHOR]

The tree crown is an important part of a tree and is closely related to forest growth status, forest canopy density, and other forest growth indicators. Chinese fir (Cunninghamia lanceolata (Lamb.) Hook) is an important tree species in southern China. A three-dimensional (3D) visualization assistant decision-making system of plantations could be improved through the construction of crown contour envelope models (CCEMs), which could aid plantation production. The goal of this study was to establish CCEMs, based on random forest and mathematical modeling, and to compare them. First, the regression equation of a tree crown was calculated using the least squares method. Then, forest characteristic factors were screened using methods based on mutual information, recursive feature elimination, least absolute shrink and selection operator, and random forest, and the random forest model was established based on the different screening results. The accuracy of the random forest model was higher than that of the mathematical modeling. The best performing model based on mathematical modeling was the quartic polynomial with the largest crown radius as the variable (R-squared (R2) = 0.8614 and root mean square error (RMSE) = 0.2657). Among the random forest regression models, the regression model constructed using mutual information as the feature screening method was the most accurate (R2 = 0.886, RMSE = 0.2406), which was two percentage points higher than mathematical modeling. Compared with mathematical modeling, the random forest model can reflect the differences among trees and aid 3D visualization of a Chinese fir plantation. [ABSTRACT FROM AUTHOR]

Forest growing stem volume (GSV) reflects the richness of forest resources as well as the quality of forest ecosystems. Remote sensing technology enables robust and efficient GSV estimation as it greatly reduces the survey time and cost while facilitating periodic monitoring. Given its red edge bands and a short revisit time period, Sentinel-2 images were selected for the GSV estimation in Wangyedian forest farm, Inner Mongolia, China. The variable combination was shown to significantly affect the accuracy of the estimation model. After extracting spectral variables, texture features, and topographic factors, a stepwise random forest (SRF) method was proposed to select variable combinations and establish random forest regressions (RFR) for GSV estimation. The linear stepwise regression (LSR), Boruta, Variable Selection Using Random Forests (VSURF), and random forest (RF) methods were then used as references for comparison with the proposed SRF for selection of predictors and GSV estimation. Combined with the observed GSV data and the Sentinel-2 images, the distributions of GSV were generated by the RFR models with the variable combinations determined by the LSR, RF, Boruta, VSURF, and SRF. The results show that the texture features of Sentinel-2's red edge bands can significantly improve the accuracy of GSV estimation. The SRF method can effectively select the optimal variable combination, and the SRF-based model results in the highest estimation accuracy with the decreases of relative root mean square error by 16.4%, 14.4%, 16.3%, and 10.6% compared with those from the LSR-, RF-, Boruta-, and VSURF-based models, respectively. The GSV distribution generated by the SRF-based model matched that of the field observations well. The results of this study are expected to provide a reference for GSV estimation of coniferous plantations. [ABSTRACT FROM AUTHOR]

Rubber (Hevea brasiliensis Muell.) plantations constitute one of the most important agro-ecosystems in the tropical region of China and Southeast Asia, playing an important role in the carbon budget there. Accurately obtaining their biomass over a large area is challenging because of difficulties in acquiring the Diameter at Breast Height (DBH) through remote sensing and the problem of biomass saturation. The stand age, which is closely related to the forest biomass, was proposed for biomass estimation in this study. A stand age map at an annual scale for Hainan Island, which is the second largest natural rubber production base in China, was generated using all Landsat and Sentinel-2 (LS2) data (1987–2017). Scatter plots and the correlation coefficient method were used to explore the relationship (e.g., biomass saturation) between rubber biomass and different LS2-based variables. Subsequently, a regression model fitted with the stand age (R2 = 0.96) and a Random Forest (RF) model parameterizing with LS2-based variables and/or the stand age were respectively employed to estimate rubber biomass for Hainan Island. The results show that rubber biomass was saturated around 65 Mg/ha with all LS2-based variables. The regression model estimated biomass accurately (R2 = 0.79 and Root Mean Square Error (RMSE) = 14.00 Mg/ha) and eliminated the saturation problem significantly. In addition to LS2-based variables, adding a stand age parameter to the RF models was found to significantly improve the prediction accuracy (R2 = 0.82–0.96 and RMSE = 4.08–10.59 Mg/ha, modeling using samples of different biomass sizes). However, all RF models overestimated the biomass of young plantations and underestimated the biomass of old plantations. A hybrid model integrating the optimal results of RF and regression models reduced estimation bias and generated the best performance (R2 = 0.83 and RMSE = 12.48 Mg/ha). The total rubber biomass of Hainan Island in 2017 was about 5.40 × 107 Mg. The northward and westward expansions after 2000 had great impact on the biomass distribution, leading to a higher biomass density for the inland coastal strip from south to northeast and a lower biomass density in the northern and western regions. [ABSTRACT FROM AUTHOR]

Understanding the spatial distribution of populations at a finer spatial scale has important value for many applications, such as disaster risk rescue operations, business decision-making, and regional planning. In this study, a random forest (RF)-based population density mapping method was proposed in order to generate high-precision population density data with a 100 m × 100 m grid in mainland China in 2015 (hereafter referred to as 'Popi'). Besides the commonly used elevation, slope, Normalized Vegetation Index (NDVI), land use/land cover, roads, and National Polar Orbiting Partnership/Visible Infrared Imaging Radiometer Suite (NPP/VIIRS), 16,101,762 records of points of interest (POIs) and 2867 county-level censuses were used in order to develop the model. Furthermore, 28,505 township-level censuses (74% of the total number of townships) were collected in order to evaluate the accuracy of the Popi product. The results showed that the utilization of multi-source data (especially the combination of POIs and NPP/VIIRS data) can effectively improve the accuracy of population mapping at a finer scale. The feature importances of the POIs and NPP/VIIRS are 0.49 and 0.14, respectively, which are higher values than those obtained for other natural factors. Compared with the Worldpop population dataset, the Popi data exhibited a higher accuracy. The number of accurately-estimated townships was 19,300 (67.7%) in the Popi product and 16,237 (56.9%) in the Worldpop product. The Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) were 14,839 and 7218, respectively, for Popi, and 18,014 and 8572, respectively, for Worldpop. The research method in this paper could provide a reference for the spatialization of other socioeconomic data (such as GDP). [ABSTRACT FROM AUTHOR]

• Establish and evaluate a predictive (random forest) model to extrapolate the distribution of soil arthropods at a large scale. • Determine the environment variables affecting the spatial distribution of soil arthropods. • Results show that the random forest model can effectively predict the spatial distribution of soil arthropods on a large scale. • The nature index of cultivated land was the most important variable for the spatial distribution of soil arthropods. • The prediction results showed that the number of soil arthropods in Changtu County China were at a low level. Soil arthropods with a body size of more than 2 mm are a significant component of terrestrial ecosystems and play an important role in biogeochemical cycles. Understanding the spatial distribution of soil arthropods is crucial for biodiversity conservation and land remediation. Studies on the spatial distribution of soil arthropods have mostly been performed at a small scale. However, due to the difficulties in obtaining soil arthropods, research on their spatial distribution patterns at a large scale is scarce and the details remain unclear. Therefore, it is important to clarify the spatial distribution pattern of soil arthropods at a large scale and the influence of environmental variables on them to develop targeted protection measures. In this study, we used a random forest model combined with environmental variables to map the spatial distribution of soil arthropods in cultivated land in Changtu County, China, and evaluated the prediction accuracy of the model. In addition, we ranked the relative importance of environmental variables to understand their influence on the spatial distribution of soil arthropods. The results showed the following: 1) The mean absolute prediction error was 13.9, the root mean square error was 21.5, R2 was 0.53, and Lin's concordance correlation coefficient of the model was 0.48, which indicated that random forest can be used to predict the spatial distribution of soil arthropods at a large scale. 2) The nature index of agricultural land, which represents the comprehensive quality of cultivated lands, and landscape division index were the two most important environmental variables affecting the spatial distribution of soil arthropods, with relative importance values of 19.3 % and 12.9 %, respectively. 3) The number of soil arthropods was low in the study area and the soil arthropods were aggregated in their distribution. We believe that accurately predicting the distribution of soil arthropods can aid in the rational distribution and management of farmland landscape and the protection of soil arthropods. [ABSTRACT FROM AUTHOR]