Search

Imagery acquired by the Moderate-resolution Imaging Spectroradiometer (MODIS) provides a global archive of dailyNormalized Difference Snow Index (NDSI) at 500 m nominal resolution since the year 2000. While Sentinel-2 (S2) NDSI provides an increased spatial resolution of 20 m since the year 2015, the temporal resolution amounts to only 5 days and thus lacks the high temporal resolution of MODIS. Efforts to combine NDSI datasets for an increased temporal and spatial resolution have so far focused on the deriving binary snow cover maps or combining data from other sensors. In contrast, we produce fine scale (20 m) fractional snow cover (FSC) by downscaling MODIS NDSI to S2 resolution. Random forest regression predicts S2 NDSI based on dynamic features (MODIS NDSI, day-of-year) and static, topographic features for an alpine study site. Subsequently, FSC is derived from S2 NDSI. Cross-validation results in R2 of 0.795 and RMSE of 0.155 for FSC and outperforms common resampling methods. Multi-annual S2 NDSI metrics are able to slightly improve model accuracy. Our results suggest that combining topographical data and low-resolution NDSI allows to produce daily, high-resolution S2 NDSI and FSC and improve fine scale characterization of snow cover dynamics in mountain landscapes. [ABSTRACT FROM AUTHOR]

Mato Grosso state is the biggest maize producer in Brazil, with the predominance of cultivation concentrated in the second harvest. Due to the need to obtain more accurate and efficient data, agricultural intelligence is adapting and embracing new technologies such as the use of satellites for remote sensing and geographic information systems. In this respect, this study aimed to map the second harvest maize cultivation areas at Canarana-MT in the crop year 2019/2020 by using geographic object-based image analysis (GEOBIA) with different spatial, spectral, and temporal resolutions. MSI/Sentinel-2, OLI/Landsat-8, MODIS-Terra and MODIS-Aqua, and PlanetScope imagery were used in this assessment. The maize crops mapping was based on cartographic basis from IBGE (Brazilian Institute of Geography and Statistics) and the Google Earth Engine (GEE), and the following steps of image filtering (gray-level co-occurrence matrix—GLCM), vegetation indices calculation, segmentation by simple non-iterative clustering (SNIC), principal component (PC) analysis, and classification by random forest (RF) algorithm, followed finally by confusion matrix analysis, kappa, overall accuracy (OA), and validation statistics. From these methods, satisfactory results were found; with OA from 86.41% to 88.65% and kappa from 81.26% and 84.61% among the imagery systems considered, the GEOBIA technique combined with the SNIC and GLCM spectral and texture feature discriminations and the RF classifier presented a mapping of the corn crop of the study area that demonstrates an improved and aided the performance of automated multispectral image classification processes. [ABSTRACT FROM AUTHOR]

Surface urban heat islands (SUHIs) can extend beyond the urban boundaries and greatly affect the thermal environment of continuous regions over an agglomeration. Traditional urban-rural dichotomy depending on the built-up and non-urban lands is challenged in characterizing regional SUHIs, such as how to accurately quantify the intensity, spatial pattern, and scales of SUHIs, which are vulnerable to SUHIs, and what the optimal scale for conducting measures to mitigate the SUHIs. We propose a machine-learning-assisted solution to address these problems based on the thermal similarity in the Yangtze River Delta of China. We first identified the regional-level SUHI zone of approximately 42,328 km2 and 38,884 km2 and the areas that have no SUHI effects from the annual cycle of land surface temperatures (LSTs) retrieved from Terra and Aqua satellites. Defining SUHI as an anomaly on background condition, random forest (RF) models were further adopted to fit the LSTs in the areas without the SUHI effects and estimate the LST background and SUHI intensity at each grid point in the SUHI zone. The RF models performed well in fitting rural LSTs with a simulation error of approximately 0.31 °C/0.44 °C for Terra/Aqua satellite data and showed a good generalization ability in estimating the urban LST background. The RF-estimated daytime Aqua/SUHI intensity peaked at approximately 6.20 °C in August, and the Terra/SUHI intensity had two peaks of approximately 3.18 and 3.81 °C in May and August, with summertime RF-estimated SUHIs being more reliable than other SUHI types owing to the smaller simulation error of less than 1.0 °C in July–September. This machine-learning-assisted solution identified an optimal SUHI scale of 30,636 km2 and a zone of approximately 23,631 km2 that is vulnerable to SUHIs, and it provided the SUHI intensity and statistical reliability for each grid point identified as being part of the SUHI. Urban planners and decision-makers can focus on the statistically reliable RF-estimated summertime intensities in SUHI zones that have an LST annual cycle similar to that of large cities in developing effective strategies for mitigating adverse SUHI effects. In addition, the selection of large cities might strongly affect the accuracy of identifying the SUHI zone, which is defined as the areas that have an LST annual cycle similar to large cities. Water bodies might reduce the RF performance in estimating the LST background over urban agglomerations. [ABSTRACT FROM AUTHOR]

Climate-driven changes in fire regimes are expected across the pan-Arctic region. Trends in arctic fires are thought to be generally increasing; however, fire mapping across the region is far from comprehensive or systematic. We developed a new detection workflow and built a dataset of unrecorded tundra fires in Canada using Landsat data. We built a reference dataset of spectral indices from previously mapped fires in northern Canada to train a Random Forest model for detecting new fires between 1986 and 2022. In addition, we used time series information for each pixel to reduce false positives and narrow the large search space down to a finite set of regions that had experienced changes. We found 209 previously undetected fires in the Arctic and sub-Arctic regions, increasing the mapped burned area by approximately 30%. The median fire size was small, with roughly 3/4 of the fires being <100 ha in size. The majority of newly detected fires (69%) did not have satellite-derived hotspots associated with them. The dataset presented here is commission error-free and can be viewed as a reference dataset for future analyses. Moreover, future improvements and updates will leverage these data to improve the detection workflow outlined here, particularly for small and low-severity fires. These data can facilitate broader analyses that examine trends and environmental drivers of fire across the Arctic region. Such analyses could begin to untangle the mechanisms driving heterogeneous fire responses to climate observed across regions of the Circumpolar North. [ABSTRACT FROM AUTHOR]

Aerosols play a crucial role in the climate system due to direct and indirect effects, such as scattering and absorbing radiant energy. They also have adverse effects on visibility and human health. Humans are exposed to fine PM2.5, which has adverse health impacts related to cardiovascular and respiratory-related diseases. Long-term trends in PM concentrations are influenced by emissions and meteorological variations, while meteorological factors primarily drive short-term variations. Factors such as vegetation cover, relative humidity, temperature, and wind speed impact the divergence in the PM2.5 concentrations on the surface. Machine learning proved to be a good predictor of air quality. This study focuses on predicting PM2.5 with these parameters as input for spatial and temporal information. The work analyzes the in situ observations for PM2.5 over Singapore for seven years (2014–2021) at five locations, and these datasets are used for spatial prediction of PM2.5. The study aims to provide a novel framework based on temporal-based prediction using Random Forest (RF), Gradient Boosting (GB) regression, and Tree-based Pipeline Optimization Tool (TP) Auto ML works based on meta-heuristic via genetic algorithm. TP produced reasonable Global Performance Index values; 7.4 was the highest GPI value in August 2016, and the lowest was −0.6 in June 2019. This indicates the positive performance of the TP model; even the negative values are less than other models, denoting less pessimistic predictions. The outcomes are explained with the eXplainable Artificial Intelligence (XAI) techniques which help to investigate the fidelity of feature importance of the machine learning models to extract information regarding the rhythmic shift of the PM2.5 pattern. [ABSTRACT FROM AUTHOR]

In the Alpine environment, snow plays a key role in many processes involving ecosystems, biogeochemical cycles, and human wellbeing. Due to the inaccessibility of mountain areas and the high spatial and temporal heterogeneity of the snowpack, satellite spatio-temporal data without gaps offer a unique opportunity to monitor snow on a fine scale. In this study, we present a random forest approach within three different workflows to combine MODIS and Sentinel-2 snow products to retrieve daily gap-free snow cover maps at 20 m resolution. The three workflows differ in terms of the type of ingested snow products and, consequently, in the type of random forest used. The required inputs are the MODIS/Terra Snow Cover Daily L3 Global dataset at 500 m and the Sentinel-2 snow dataset at 20 m, automatically retrieved through the recently developed revised-Let It Snow workflow, from which the selected inputs are, alternatively, the Snow Cover Extent (SCE) map or the Normalized Difference Snow Index (NDSI) map, and a Digital Elevation Model (DEM) of consistent resolution with Sentinel-2 imagery. The algorithm is based on two steps, the first to fill the gaps of the MODIS snow dataset and the second to downscale the data and obtain the high resolution daily snow time series. The workflow is applied to a case study in Gran Paradiso National Park. The proposed study represents a first attempt to use the revised-Let It Snow with the purpose of extracting temporal parameters of snow. The validation was achieved by comparison with both an independent dataset of Sentinel-2 to assess the spatial accuracy, including the snowline elevation prediction, and the algorithm's performance through the different topographic conditions, and with in-situ data collected by meteorological stations, to assess temporal accuracy, with a focus on seasonal snow phenology parameters. Results show that all of the approaches provide robust time series (overall accuracies of A1 = 93.4%, and A2 and A3 = 92.6% against Sentinel-2, and A1 = 93.1%, A2 = 93.7%, and A3 = 93.6% against weather stations), but the first approach requires about one fifth of the computational resources needed for the other two. The proposed workflow is fully automatic and requires input data that are readily and globally available, and promises to be easily reproducible in other study areas to obtain high-resolution daily time series, which is crucial for understanding snow-driven processes at a fine scale, such as vegetation dynamics after snowmelt. [ABSTRACT FROM AUTHOR]

Land-use and land-cover (LULC) mapping in the complex area is a challenging task due to the mixed vegetation patterns, and rough mountains with fast-flowing rivers. In Vietnam, LULC update is not frequently. In this study, we applied a supervised machine learning (Random forest—RF) approach to mapping LULC in Thanh Hoa province, Vietnam from 2011 to 2015 utilizing multitemporal Normalized Difference Vegetation Index (NDVI) data from MODIS, combined with topographic features. Random forest classification (RFC) reached a total prediction accuracy of 91% and Kappa coefficient (K) of 0.89 across eight LULCs. Besides, the results showed that the features extracted from time-series NDVI comprising the mean of yearly NDVI, the sum of NDVI, and the topography were the important variables controlling the LULC classification. For similar studies on the distribution of LULC, the method proposed in this study could be helpful. [ABSTRACT FROM AUTHOR]

The objective of this paper is to contribute to improve ETa estimation in semiarid environments by proposing two variations to the TDMT model. These variations are based on the use of MODIS products from TERRA or AQUA satellites and on the use of NDVI instead of EVI, to estimate the fraction of vegetation cover. The proposed changes were validated with the original methodology for the 2012-2014 period with data obtained from two flux towers (ES-LJu and ES-Amo). The best results were obtained when using the alternative methodology (RMSE of 0.64-0.67mm in ES-Lju and of 0.97-1.02mm in ES-Amo). Additionally, a correction of the temporal systematic errors of the model based on RandomForest is proposed. With this correction, RMSE of 0.31-0.35mm in ES-Lju and 0.30-0.34mm in ES-Amo were obtained. The spatial distribution obtained with this corrected model is the most consistent with the characteristics of the study area. [ABSTRACT FROM AUTHOR]

The MODIS 8-day composite evapotranspiration (ET) product (MOD16A2) is widely used to study large-scale hydrological cycle and energy budgets. However, the MOD16A2 spatial resolution (500 m) is too coarse for local and regional water resource management in agricultural applications. In this study, we propose a Deep Neural Network (DNN)-based MOD16A2 downscaling approach to generate 30 m ET using Landsat 8 surface reflectance and temperature and AgERA5 meteorological variables. The model was trained at a 500 m resolution using the MOD16A2 ET as reference and applied to the Landsat 8 30 m resolution. The approach was tested on 15 Landsat 8 images over three agricultural study sites in the United States and compared with the classical random forest regression model that has been often used for ET downscaling. All evaluation sample sets applied to the DNN regression model had higher R2 and lower root-mean-square deviations (RMSD) and relative RMSD (rRMSD) (the average values: 0.67, 2.63 mm/8d and 14.25%, respectively) than the random forest model (0.64, 2.76 mm/8d and 14.92%, respectively). Spatial improvement was visually evident both in the DNN and the random forest downscaled 30 m ET maps compared with the 500 m MOD16A2, while the DNN-downscaled ET appeared more consistent with land surface cover variations. Comparison with the in situ ET measurements (AmeriFlux) showed that the DNN-downscaled ET had better accuracy, with R2 of 0.73, RMSD of 5.99 mm/8d and rRMSD of 48.65%, than the MOD16A2 ET (0.65, 7.18 and 50.42%, respectively). [ABSTRACT FROM AUTHOR]

Extreme forest fires have been a historic concern in the forests of Canada, the Russian Federation, and the USA, and are now an increasing threat in boreal Europe, where recent fire events in 2014 and 2018 drew attention to Sweden. Our study objective was to understand the vulnerability of Swedish forests to fire by spatially analyzing historical burned areas, and to link fire events with weather, landscape, and fire-related socioeconomic factors. We developed an extensive database of 1 × 1 km2 homogenous grids, where monthly burned areas were derived from the MODIS FireCCI51 dataset. The database consists of various socio-economic, topographic-, forest-, and weather-related remote sensing products. To include new factors in the IIASA's FLAM model, we developed a random forest model to assess the spatial probabilities of burned areas. Due to Sweden's geographical diversity, fire dynamics vary between six biogeographical zones. Therefore, the model was applied to each zone separately. As an outcome, we obtained probabilities of burned areas in the forests across Sweden and observed burned areas were well captured by the model. The result accuracy differs with respect to zone; the area under the curve (AUC) was 0.875 and 0.94 for zones with few fires, but above 0.95 for zones with a higher number of fire events. Feature importance analysis and their variability across Sweden provide valuable information to understand the reasons behind forest fires. The Fine Fuel Moisture Code, population and road densities, slope and aspect, and forest stand volume were found to be among the key fire-related factors in Sweden. Our modeling approach can be extended to hotspot mapping in other boreal regions and thus is highly policy-relevant. Visualization of our results is available in the Google Earth Engine Application. [ABSTRACT FROM AUTHOR]

The aboveground biomass (AGB) of a forest is an important indicator of the forest's terrestrial carbon storage and its relation to climate change. Due to the advantage of extensive spatial coverage and low cost, coarse-resolution remote sensing data is the main data source for wall-to-wall mapping of forest AGB at the regional scale. Despite this, improving the accuracy and efficiency of forest AGB estimation is a major challenge. In this study, two optical imageries, Moderate Resolution Imaging Spectroradiometer (MODIS) 500 m imagery and Fengyun-3C Visible and Infrared Radiometer (FY-3C VIRR) 1000 m imagery, were used and compared for forest AGB estimation in Yunnan Province, southwest China. One parametric approach, multiple linear regression (MLR), and two nonparametric approaches, k-nearest neighbor (KNN) and random forest (RF), were applied for the two imagery datasets, respectively. We evaluated the performance of the combination of remote sensing data and modeling approaches by comparing the accuracies and also explored the potential of FY-3C imagery data in forest AGB estimation at the regional scale as it was used for this purpose for the first time. We found that the machine learning models KNN and RF provided better results than MLR. From the three approaches for both MODIS and FY-3C imagery, RF performed best with R2 values of 0.84 and 0.81 and RMSE of 23.18 and 23.43, respectively. Estimation of forest AGB based on MODIS was marginally better than the estimation based on FY-3C. FY-3C imagery could therefore be an additional optical remote sensing data source of coarse spatial resolution, comparable to MODIS data which has been widely used for regional forest AGB estimation. Indices related to forest canopy moisture levels from both types of imagery were sensitive to forest AGB. The RF model and MODIS imagery were then applied to map the spatial variation of forest AGB of Yunnan Province. As a result of our study, we determined that Yunnan Province has a total forest AGB of 2123.22 Mt, with a mean value of 58.05 t/ha for forestland in 2016. [ABSTRACT FROM AUTHOR]

With the growth of cloud computing, the use of the Google Earth Engine (GEE) platform to conduct research on water inversion, natural disaster monitoring, and land use change using long time series of Landsat images has also gradually become mainstream. Landsat images are currently one of the most important image data sources for remote sensing inversion. As a result of changes in time and weather conditions in single-view images, varying image radiances are acquired; hence, using a monthly or annual time scale to mosaic multi-view images results in strip color variation. In this study, the NDWI and MNDWI within 50 km of the coastline of the Yucatán Peninsula from 1993 to 2021 are used as the object of study on GEE platform, and mosaic areas with chromatic aberrations are reconstructed using Landsat TOA (top of atmosphere reflectance) and SR (surface reflectance) images as the study data. The DN (digital number) values and probability distributions of the reference image and the image to be restored are classified and counted independently using the random forest algorithm, and the classification results of the reference image are mapped to the area of the image to be restored in a histogram-matching manner. MODIS and Sentinel-2 NDWI products are used for comparison and validation. The results demonstrate that the restored Landsat NDWI and MNDWI images do not exhibit obvious band chromatic aberration, and the image stacking is smoother; the Landsat TOA images provide improved results for the study of water bodies, and the correlation between the restored Landsat SR and TOA images with the Sentinel-2 data is as high as 0.5358 and 0.5269, respectively. In addition, none of the existing Landsat NDWI products in the GEE platform can effectively eliminate the chromatic aberration of image bands. [ABSTRACT FROM AUTHOR]

The timely and accurate estimation of grassland aboveground biomass (AGB) is important. Machine learning (ML) has been widely used in the past few decades to deal with complex relationships. In this study, based on an 11-year period (2005–2015) of AGB data (1620 valid AGB measurements) on the Three-River Headwaters Region (TRHR), combined with remote sensing data, weather data, terrain data, and soil data, we compared the predictive performance of a linear statistical method, machine learning (ML) methods, and evaluated their temporal and spatial scalability. The results show that machine learning can predict grassland biomass well, and the existence of an independent validation set can help us better understand the prediction performance of the model. Our findings show the following: (1) The random forest (RF) based on variables obtained through stepwise regression analysis (SRA) was the best model (R2vad = 0.60, RMSEvad = 1245.85 kg DW (dry matter weight)/ha, AIC = 5583.51, and BIC = 5631.10). It also had the best predictive capability of years with unknown areas (R2indep = 0.50, RMSEindep = 1332.59 kg DW/ha). (2) Variable screening improved the accuracy of all of the models. (3) All models' predictive accuracy varied between 0.45 and 0.60, and the RMSE values were lower than 1457.26 kg DW/ha, indicating that the results were reliably accurate. [ABSTRACT FROM AUTHOR]

Remotely sensed vegetation indices have been widely used to estimate live fuel moisture content (LFMC). However, marked differences in vegetation structure affect the relationship between field-measured LFMC and reflectance, which limits spatial extrapolation of these indices. To overcome this limitation, we explored the potential of random forests (RF) to estimate LFMC at the subcontinental scale in the Mediterranean basin wildland. We built RF models (LFMCRF) using a combination of MODIS spectral bands, vegetation indices, surface temperature, and the day of year as predictors. We used the Globe-LFMC and the Catalan LFMC monitoring program databases as ground-truth samples (10,374 samples). LFMCRF was calibrated with samples collected between 2000 and 2014 and validated with samples from 2015 to 2019, with overall root mean square errors (RMSE) of 19.9% and 16.4%, respectively, which were lower than current approaches based on radiative transfer models (RMSE ~74–78%). We used our approach to generate a public database with weekly LFMC maps across the Mediterranean basin. [ABSTRACT FROM AUTHOR]

• Random forests were used to estimate fractional snow cover (FSC) based on 14 variables. • Landsat and MODIS data were integrated to improve snow cover estimation. • The full and semi models showed similarity, trimmed-model performed weaker. • The model's performance was higher in grassland regions than in forested regions. • Full and semi models achieved the highest R-values in the middle elevation zone. Accurate estimation of snow-covered area (SCA) is vital for effective water resource management, especially in snowmelt-dependent regions like the Kabul River Basin (KRB). It serves as a reference point for comparing expected variations in water availability driven by climate change, particularly in arid and semi-arid regions like the KRB. In this study, fractional snow cover (FSC) was estimated across the KRB using Landsat and moderate resolution imaging spectroradiometer (MODIS) datasets. For this purpose, 34 Landsat-8 and MODIS image-pairs were acquired covering the snowfall period from 01 October 2018 and 31 March 2021. The training dataset consisted of 31 image-pairs, while the remaining three were used as an independent test dataset. Sample sizes and training strategies (i.e., full, semi, and trimmed-models), three each, were evaluated to understand the relevance of the predictor variables. The full-model incorporated MODIS surface reflectance bands (SRB) 1–7 spectral indices, topography and landcover while the semi-model included MODIS SRB 1–7 and indices. The trimmed-model only utilized SRB 1–7. Random Forests (RF) facilitated FSC mapping with Landsat-8 data as a reference. The findings indicated comparable performance between full and semi models, whereas the trimmed-model exhibited weaker performance. The correlation coefficient (R) of the full, semi and trimmed models ranged from 0.83 to 0.92, 0.83–0.92 and 0.82–0.87 respectively. The models performed strongly in grassland regions (R = 0.89–0.90), but moderately in forested areas (R = 0.43–0.53). This approach results in improved MODIS-based SCA-mapping in the Hindukush Mountains, facilitating better water resource management in the region. [ABSTRACT FROM AUTHOR]

Surface longwave net radiation (SLNR) is the key component of the surface radiation budget. SLNR is estimated by using the conventional method, in which surface longwave downward radiation (SLDR) and surface longwave upward radiation (SLUR) are estimated separately, and then SLNR is calculated by subtracting SLUR from SLDR. However, uncertainties in SLDR and SLUR may accumulate and propagate into the calculated SLNR, affecting the SLNR estimate accuracy. This paper proposes a direct algorithm that links clear-sky SLNR with the top-of-atmosphere (TOA) radiances in MODIS channels 29, 31 and 32 and column water vapour (CWV). The multivariate regression (MVR) and extreme gradient boosting (×GBoost) were employed to establish the relationships sequentially with collocated ground measurements and satellite data for high-latitude, mid-latitude, and low-latitude areas. According to the training and validation results, the GBoost model performs slightly better than the MVR model, which indicates that weak nonlinearity exists between clear-sky SLNR and its predictors. The absolute biases of the XGBoost models are approximately 1.0 Wm−2, and the RMSEs of the XGBoost model are 17.77, 27.33, and 19.60 Wm−2 in high-latitude, mid-latitude, and low-latitude areas. These results indicate that nonlinear relationships exist between clear-sky SLNR and predictors, but this nonlinearity is not strong. The superiority of the direct algorithm over the conventional method was also demonstrated. In the future, we will collect more data to improve the performance of the proposed direct algorithm especially in high-latitude and mid-latitude areas. [ABSTRACT FROM AUTHOR]

Grassland habitats provide many ecosystem services but are threatened by agricultural intensification and urbanization. While the lack of accurate and comprehensive inventories at the national scale makes them difficult to manage, advances in spatial modeling using open remote sensing data and open-source software, as well as the increasing use of ecological archives, provide new perspectives for mapping natural habitats. In this context, this study evaluated the contribution of spectral and environmental variables to discriminate and then map grassland habitats throughout France. To this end, 92 spectral variables derived from moderate-resolution imaging spectroradiometer data, 19 bioclimatic variables derived from WorldClim data, 4 topographic variables derived from the European Union Digital Elevation Model (DEM), and 8 soil variables derived from SoilGrids data were combined at a spatial resolution of 250 m. Reference plots that characterized 6 and 21 grassland ecosystems at European Nature Information System (EUNIS) levels 2 (broad habitats) and 3 (habitats), respectively, were collected from vegetation archives. We first performed descriptive analysis that included habitat description, ordination, and pairwise separability. We then performed predictive analysis of grassland habitats using a cross-validated random forest model that included a spatial constraint. While environmental and spectral variables characterized most grassland habitats well and consistently, some confusion occurred between habitats with similar abiotic conditions. The main grassland habitat types were correctly mapped at EUNIS level 2 (F 1 score = 0.68), but not at EUNIS level 3 (F 1 score = 0.52). In addition, the two variables that contributed most to the model were the near-infrared spectral band in spring and the minimum temperature of the coldest month. The model's prediction at EUNIS level 2 for mainland France provides the map of grassland habitats at a new spatial scale. [ABSTRACT FROM AUTHOR]

Ice plays key roles in regulating hydrological, ecological, biogeochemical, and socioeconomic functions of lakes. Long-term in situ lake ice phenological records indicate that lake ice is trending toward later freeze-up, earlier breakup, and a shorter ice duration. Parallel to study of lake ice using in situ records and process-based models, satellite remote sensing can expand our understanding of lake ice change over large spatial scales. However, most remote sensing studies have focused on large lakes or short periods of time, which may not robustly represent changes over multidecadal time periods or in the much more numerous small lakes. Here, we present a random forest model, Sensitive Lake Ice Detection (SLIDE), to accurately extract ice conditions from Landsat TM, ETM+, and OLI images. We trained the model using a manually labeled lake ice dataset (1089 labeled areas over 995 lakes globally). Our results show that our model achieves accurate classification between ice/snow and water (accuracy: 97.8%, kappa coefficient: 95.5%). Comparing Landsat-derived ice cover with in situ ice conditions, we show that our model produces less bias, lower RMSE, and higher kappa than does the Landsat snow/ice flag from the quality assessment band. This is especially true during the transitional period surrounding the ice on and off dates reported from in situ (mean bias −7.3% from our model, −17.3% from the Landsat quality band). Our results demonstrate the feasibility of mining the rich Landsat archive to study lake ice dynamics and of better flagging ice-affected lake observations. [ABSTRACT FROM AUTHOR]

The occurrence of drought is related to complicated interactions between many factors, such as precipitation, temperature, evapotranspiration and vegetation. In this study, the relationships between drought and precipitation, temperature, vegetation and evapotranspiration were investigated with a random forest (RF), and a new combined drought monitoring index (CDMI) was constructed. The effectiveness of the CDMI in monitoring drought in Shaanxi Province was verified by the in situ 1 ∼ 12-month standardized precipitation index (SPI); relative soil moisture (RSM) and four other commonly used remote sensing drought monitoring indices. The results show that CDMI is more correlated with the SPI and RSM than the four indices. Moreover, the spatial distributions of drought for the CDMI and RSM are similar. Therefore, the CDMI can be used to monitor droughts in Shaanxi Province, and machine learning can explore the relationships between various factors and establish a drought index without knowledge of the causal mechanisms of these factors. [ABSTRACT FROM AUTHOR]

Lakes in Central Asia, characterized largely by a semiarid climate and delicate ecosystems, are quite sensitive and vulnerable to climate change but remain less exploited, due mostly to a lack of in-situ measurements. Lake ice phenology is an essential climate variable identified by the Global Climate Observation System. To overcome the overestimation of clouds by the MODIS cloud mask, we developed and distributed a two-step Random Forest algorithm based on MODIS data on the Google Earth Engine. This algorithm first distinguishes cloudy pixels from non-cloudy pixels, and then ice pixels are discriminated from non-cloudy pixels using additional thermal infrared information. We validated the algorithm by implementing it over several boreal lakes with in-situ records in North America and compared our results with existing lake ice phenology datasets. We subsequently retrieved lake ice phenology for 20 large lakes (> 100 km2) in Central Asia from 2001 to 2021 and examined trends in lake ice phenology over this period. On average, lakes in Central Asia underwent later freezing at 0.13 d/yr, earlier thawing at 0.32 d/yr, and shorter ice duration at 0.45 d/yr, with over half of the lakes witnessing significant trends in at least one ice phenology indicator. To evaluate the impact of meteorological factors on lake ice phenology, we built a linear mixed model that predicts ice phenology using meteorological variables from ERA5. Daily mean air temperature dominates ice phenology, with mean air temperature during October to June in the following year explaining approximately 35% of the variance of the duration between freezing and thawing for each lake. Our study offers a comprehensive analysis of lake ice phenology in Central Asia under climate change. • A two-step ice classification approach addresses overestimation of MODIS cloud mask. • A complete and consistent 21-year ice phenology dataset for 20 Central Asia lakes. • Over half of the lakes have significant trends in at least one phenology indicator. • Mean air temperature explains approximately 35% of the variance of ice duration. [ABSTRACT FROM AUTHOR]

As a common phenomenon over Antarctica, blowing snow (BLSN), especially the large BLSN storms, play an important role in the Antarctic surface mass balance, radiation budget, and planetary boundary layer processes. This study presents the work on BLSN storm identification and analysis with observations from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the Aqua satellite. Spectral analysis shows that BLSN identification is feasible with MODIS daytime data. A random forest machine learning model is developed and observations from the Cloud‐Aerosol Lidar with Orthogonal Polarization are used for training. Model performance results show that machine‐learning based classification can achieve over 90% overall accuracy when classifying MODIS pixels into cloud, clear, and BLSN categories. The machine learning model is applied to MODIS observations during the month of October 2009 for BLSN storm analysis. Results show that the size of BLSN storms has a large spectrum and can reach hundreds of thousands km2. The MODIS based BLSN storm frequency map extends the Cloud‐Aerosol Lidar and Infrared Pathfinder Satellite Observations coverage limit from 82°S to the South Pole. A BLSN storm belt, which extends from the South Pole region to the coastal area between 130°E and 160°E along the Transantarctic Mountains, provides a potential pathway of snow transport. These results are important in improving the understanding of BLSN impact on Antarctic surface mass balance and boundary layer processes. [ABSTRACT FROM AUTHOR]

Rising global temperatures over the last decades have increased heat exposure among populations worldwide. An accurate estimate of the resulting impacts on human health demands temporally explicit and spatially resolved monitoring of near‐surface air temperature (Ta). Neither ground‐based nor satellite‐borne observations can achieve this individually, but the combination of the two provides synergistic opportunities. In this study, we propose a two‐stage machine learning‐based hybrid model to estimate 1 × 1 km2 gridded intra‐daily Ta from surface skin temperature (Ts) across the complex terrain of Israel during 2004–2016. We first applied a random forest (RF) regression model to impute missing Ts from the Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua and Terra satellites, integrating Ts from the geostationary Spinning Enhanced Visible and InfraRed Imager (SEVIRI) satellite and synoptic variables from European Centre for Medium‐Range Weather Forecasts' (ECMWF) ERA5 reanalysis data sets. The imputed Ts are in turn fed into the Stage 2 RF‐based model to estimate Ta at the satellite overpass hours of each day. We evaluated the model's performance applying out‐of‐sample fivefold cross validation. Both stages of the hybrid model perform very well with out‐of‐sample fivefold cross validated R2 of 0.99 and 0.96, MAE of 0.42°C and 1.12°C, and RMSE of 0.65°C and 1.58°C (Stage 1: imputation of Ts, and Stage 2: estimation of Ta from Ts, respectively). The newly proposed model provides excellent computationally efficient estimation of near‐surface air temperature at high resolution in both space and time, which helps further minimize exposure misclassification in epidemiological studies. [ABSTRACT FROM AUTHOR]

Soil moisture (SM) is an important index of soil drought, and it directly controls the energy balance and water cycle of the land surface. As an indicator and amplifier of global warming, the Tibetan Plateau (TP) is becoming warmer and wetter. Because of its particular geographical environment, large‐scale measurements of SM on the TP can only be achieved by satellite remote sensing. The resolution of current SM product of the Soil Moisture Active Passive (SMAP) satellite is 36 km, which is insufficient for many practical applications. In this study, the Moderate Resolution Imaging Spectroradiometer (MODIS) and Digital Elevation Model (DEM) are applied to increase the resolution of SM down to 1 km using the Random Forest (RF) algorithm. The preliminary results of the proposed algorithm are evaluated by station observations and other reanalysis products. The downscaled results are more consistent with the in situ observations, the Land Data Assimilation System (CLDAS) from China Meteorological Administration (CMA), and the Global Land Data Assimilation System (GLDAS) from National Aeronautics and Space Administration (NASA) than the original SMAP product. The downscaling algorithm is most effective for grasslands. It is demonstrated that high‐resolution SM products can be generated by fusing various features using machine‐learning algorithms. Key Points: A machine‐learning algorithm is proposed to increase the resolution of soil moisture products of remote sensing on the Tibetan PlateauTopographical features are applied to estimate soil moisture in complex terrains, so that the accuracy of estimation could be improvedRandom forest algorithm is applied to combine various features, and the results are expected to well fit the real situation [ABSTRACT FROM AUTHOR]

Satellite-based remote sensing is a powerful form of technology that can provide food security policy makers with reliable information. This information allows them to estimate final crop yields on a global scale within reasonable time frames and with higher spatial resolution than with the use of pure statistical data. Satellite-based crop yield estimation methods are commonly based on the high correlation between the crop yield and the vegetation index (VI), taken at a specific phenological stage. Although VI-based crop yield estimation methods that make use of one approximation formula can easily and effectively estimate the spatial distribution of corn and soybean yields in the United States, there are still some associated drawbacks to this approach that result in the underestimation of crop yields, especially in irrigated regions. Furthermore, a fundamental problem with this approach is the difficulty in evaluation of environmental stress-related physiological disorders such as sterility, which cannot be evaluated based on VIs as an alternative value to biomass. This study's objective was, thus, to overcome the limitations associated with the conventional approach by incorporating additional environmental variables into the proposed method along with the application of a random forest regression algorithm for estimating United States (US) corn and soybean yields with higher accuracy. This study compared three methods: (1) a conventional method based on a linear regression model (LM method) calibrated using limited past data, (2) a method, which was slightly altered from the LM method in terms of the use of a polynomial regression model (PM method), and (3) the newly proposed method, which involved the application of a random forest regression algorithm and the use of irrigated harvested cropland percentage and reanalysis data for temperature, precipitation, shortwave radiation, and soil moisture (RF method). The time-series correlation between the moderate resolution imaging spectroradiometer (MODIS) wide dynamic ranged vegetation index (WDRVI) and corn and soybean yields were analyzed as part of a preliminary investigation to determine the best time for recording the MODIS WDRVI as an explanatory variable in the study area. The results revealed that the MODIS WDRVI demonstrated the highest correlation with county-level statistical yields 13 days before the silking stage for corn and 6 days before the setting pods stage for soybeans. The regression formulas for the LM and the PM method were developed based on assigning the MODIS WDRVI to these phenological stages as explanatory variables. The advantage of the PM method over the LM method was found to be its adaptability to high-yield counties because of the inherent effect of using a polynomial regression equation. The LM method, which made use of a linear regression equation calibrated using limited past data (2009–2010), could not be adapted to increased yields encountered in recent years without recalibration with the latest data. The RF method learning models were individually optimized for each state and crop. This optimization revealed, that our learning model that incorporated every available variable did not always perform best, probably due to overfitting. In the major irrigated states of Kansas, and Nebraska, the spatial data of the percentage of irrigated harvested cropland improve the estimation accuracy of the RF method for both corn and soybean. In the states of Illinois and Iowa, the RF method, which incorporated primarily the weather-related variables of soil moisture, precipitation, temperature, and shortwave radiation, improved the estimation accuracy due to a response of rainfed agriculture to environmental stress. This is especially true for soybean. The validation results indicated that the estimation accuracy of the RF method (root mean square error RMSE: 0.539 t/ha for corn, 0.206 t/ha for soybeans) was higher than that of the PM method (RMSE: 0.897 t/ha for corn, 0.283 t/ha for soybeans) at the state level, particularly due to the effect of bias correction in irrigated regions. Moreover, it was confirmed that the RF method could provide an accurate estimate of the yield reduction in soybeans caused by a drought occurring during late vegetative stages in Illinois in 2003. The PM method could not be used to evaluate this drought-induced yield reduction based on the VIs and key phenological stages. According to visual depictions of corn- and soybean-yield estimation maps, the RF-derived maps corresponded better with the yield maps derived from the National Agricultural Statistics Service (NASS)-statistical data than the PM-derived maps, especially in low-yield and irrigated regions. [ABSTRACT FROM AUTHOR]

[Display omitted] • A selection strategy for base-period images of spatio-temporal fusion was proposed. • The spatio-temporal inconsistency between field samples and images was reduced. • The estimation accuracy of grassland aboveground biomass was improved. The accurate estimation of the aboveground biomass (AGB) is crucial for the sustainable utilization and management of grassland resources. Spatio-temporal inconsistencies between field samples and satellite images are major source of error in the estimation of grassland AGB. To solve this problem, this study selected the Three Rivers Headwater Region as the study area and proposed a selection strategy for base-period images of spatio-temporal fusion that was suitable for use at a large spatial scale in which cloud-free images are difficult to obtain. The spatial and temporal adaptive reflectance fusion model (STARFM) based on the selection strategy was used to generate a synthetic Normalized Difference Vegetation Index (NDVI) dataset with high spatial-temporal resolution by using the maximum value composite of GF-1 NDVI and MODIS NDVI to enhance the spatial-temporal quality of the images for field-scale application. Three estimation models for grassland AGB were then constructed by the random forest algorithm using synthetic NDVI, MODIS NDVI, GF-1 NDVI respectively, together with ancillary data. Following this, the estimation model with the highest accuracy was used to generate a 16-m eight-day time-series AGB in the growing season. The results showed: (1) The synthetic NDVI was correlated closely with the observed GF-1 NDVI, with an average R of 0.825 and a RMSE of 0.087. The temporal trend of the synthetic NDVI for each grassland type was highly consistent with that of the MODIS NDVI in the growing season with a correlation higher than 0.9. (2) The synthetic NDVI reduced the spatial difference between field samples and images to 16-fold, and the temporal difference was controlled to within four days under ideal conditions. (3) The synthetic NDVI improved the estimation accuracy of grassland AGB by about 15.9% and 19.7% (R2), and 13.7% and 17.5% (RMSE) relative to MODIS NDVI and GF-1 NDVI, respectively. (4) The time-series AGB revealed accurately the spatial distribution of and seasonal temporal variations in the grassland biomass. The results of this study may serve as scientific guidance for timely monitoring of grassland conditions and precise management of grassland resources in China. [ABSTRACT FROM AUTHOR]

The land surface's thermal dynamics follows annual and diurnal cycles that are to a large extent controlled by solar geometry. Therefore, annual and diurnal variations of land surface temperature (LST) can be modelled with relatively simple functions controlled by a small number of parameters, typically from three to six. The parameter values of the models can be determined by fitting the respective functions to time series of LST observations. Commonly either annual or diurnal LST variations are modelled or they are modelled sequentially in a two-step process. Here, we combine an annual temperature cycle (ATC) model controlled by the solar zenith angle (ATC sza) with a four-parameter version of the diurnal temperature cycle (DTC) model GOT09: this yields a new annual-diurnal temperature cycle (ADTC) model that simultaneously describes the annual and diurnal surface temperature dynamics. The proposed ADTC model is controlled by physically meaningful parameters: annual minimum temperature, annual temperature amplitude, annual maximum of daily temperature amplitude, mean time of thermal noon and time lag of maximum temperature with respect to summer solstice. Thus, the entire annual-diurnal LST dynamics is described with only five parameters. The new model was tested by fitting it to one year of LST observations obtained for five globally representative tiles of the Moderate Resolution Imaging Spectroradiometer (MODIS), onboard EOS – TERRA and EOS – AQUA satellites. For these tiles, the mean of the root mean square error (RMSE) was 4.2 K. ADTC modelled LSTs were also compared against those obtained with the standard ATC model for the four MODIS overpass times at five representative sites: for these, an overall RMSE of 1.2 K between the two models was obtained. The ADTC derived LSTs were validated against in-situ measurements from three different sites, which yielded an overall RMSE of 3.4 K. Additional investigations over five areas with different land covers (i.e. urban, lake, forest, mountain area and desert) revealed the potential of the ADTC parameters to describe the corresponding surface and climate properties. Since it is driven by solar geometry, the ADTC model reproduces double LST peaks in the tropics naturally. Furthermore, all available observations are modelled simultaneously, which means that a single set of parameters is obtained for each pixel and year. • A combined annual-diurnal temperature cycle (ADTC) model is presented. • The ADTC model simultaneously describes the annual and diurnal LST dynamics. • The model is controlled by five parameters (ADCPs) with physical meaning. • Five MODIS LST tiles were modelled with an average RMSE of 4.2 K. • The RMSE with respect to the standard ATC (Bechtel, 2012) was 1.2 K. [ABSTRACT FROM AUTHOR]

Abstract Remote sensing is an effective means of observing and detecting global aerosol distribution and changes over time, which impact human health and climate change. However, aerosol optical depth (AOD) always has low spatial coverage, which not only affects the analysis of AOD but also harms many relevant applications of the data, such as utilization to estimate PM 2.5. In our study, we utilize the random forest model, which is an effective ensemble learning method, to estimate the gaps of Moderate Resolution Imaging Spectroradiometer (MODIS) AOD data with a spatial resolution of 0.01° × 0.01° in a typical contaminated region of Beijing-Tianjin-Hebei during 2010–2016. Our model performs accurately in that the results of R2 testing exceed 0.9 and the final estimated AOD coverage achieves 100%. The average value of the AOD is 0.44 (0.41–0.47 by year) over the study period. The simulation values of AOD have an obvious seasonal distribution, with the highest AOD in summer. The AOD estimations in the southern region are higher than those in the northern region. Aerosol Robotic Network (AERONET) AOD observations are compared with MODIS AOD (R2 = 0.44) and AOD estimations (R2 = 0.36). We analyze and screen each of the variables to compute their contributions. Specifically, the elevation and 2-m dew point are the most important in modeling the AOD, while road data, snowfall depth and snowfall have the least impact on modeling the AOD. Practical applications of AOD data include estimating the various impacts of PM 2.5 concentrations on health based on the AOD observations in China's typically polluted areas that have cloud influence. We compare two measurement ranges that will most accurately model and fill the AOD data missing in areas. After careful consideration, we determine that our preferred range is 0–2. Graphical abstract Image 1 Highlights • We produced a full-coverage daily AOD based on the random forest approach. • Our models achieved high performance over the study period. • Our research is based on multi-parameter modeling. • Explore the appropriate range of AOD observation to model. [ABSTRACT FROM AUTHOR]

The aim of this paper is to map the aboveground biomass (AGB) in Gabon. First, a random forest (RF) model that relates reference AGB values to remote sensing (RS)-derived variables (mainly radar and optical images) was built, and the significant predictive variables were determined. Second, the built RF model was applied to the significant RS-derived variables to predict AGB across Gabon. The results showed that the overall RMSE (Root Mean Square Error) on the RS-derived AGB map with a spatial resolution of 50 m was 63.3 t/ha (R 2 = 0.53). To improve the accuracy of the RS-derived AGB map, the integration of LiDAR data provided by the Geoscience Laser Altimeter System (GLAS) onboard the Ice Cloud and Land Elevation Satellite (ICESat) was investigated. First, an RF model that relates reference AGB values to GLAS-derived metrics and a DEM (Digital Elevation Model) was built. Second, the calibrated RF model was applied to obtain a spatially distributed estimation of AGB (GLAS footprints geolocation) covering forested areas in Gabon, with a density of 0.13 footprints/km2. Third, the semivariogram of residuals (RS-derived AGB map – GLAS-derived AGB "surrogate AGB") was computed. Later, a regression kriging interpolation was performed by taking into account the spatial structure of residuals to provide a continuous residual map. Finally, the RS-derived AGB map and the residual map were summed, and a final AGB map was obtained. The results showed that the integration of GLAS surrogate AGB data slightly improves the accuracy of the RS-derived AGB map only for AGB values lower than 100 t/ha (bias and RMSE reduced by 13.9 and 10 t/ha, respectively). [ABSTRACT FROM AUTHOR]

Abstract Land surface temperature (LST) plays a key role in connecting land surface energy and water exchanges with near-surface atmosphere. However, the spatial distribution of LST over mountainous areas is not only strongly affected by the differences in surface thermal properties but also by the differences in thermal or radiative environment induced by the topographic variations, presenting significant terrain effect. This effect greatly hinders researches on surface energy fluxes and soil moisture estimation in these regions. To normalize the terrain effect, a random forest (RF)-based normalization approach was developed in this study and applied to Moderate Resolution Imaging Spectroradiometer (MODIS) data over a typical mountain region in the eastern part of the Tibetan Plateau, China. An LST linking model was first constructed to express the complicated interrelationship between LST and other surface variables, including normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), leaf area index (LAI), surface albedo (ALB), cumulative incident solar radiation (CSR), normalized difference water index (NDWI), surface elevation (ELV), and surface slope (SLP). The estimation results well indicated the good accuracy of the model with the coefficient of determination (R2) above 0.92 for 90 selected days in 2015. Based on the good LST linking model, the LST normalization was achieved by replacing the topography affected factors (CSR, ELV, and SLP) with reference values for each pixel when keeping the rest status factors as their original values. The normalization result was indirectly validated by the normalization results with the traditional method based on the linear correction upon the regression between surface elevation and LST. The cross-validation clearly indicated the proposed method had an obvious advantage in reducing the topography-induced LST difference based on the average decrease in the temperature range (9.51 K) and the standard deviation (1.09 K) for the images on the 90 selected days over the study area. In contrast, the traditional method was hard to capture the terrain effect only based on the relationship between LST and surface elevation due to the complex interaction between LST and other factors. Overall, the proposed method shows a good application potential for normalizing the terrain effect on LST, and the results will be helpful for the LST-based estimation of surface energy fluxes or soil moisture over mountainous areas. Highlights • A random forest-based LST terrain effect normalization method is proposed. • The method well depicts the nonlinear relationship between LST and other variables. • The normalized LST has a significant reduction in topography-induced difference. • Allowed improved estimation of LST-related variables over the mountainous areas. [ABSTRACT FROM AUTHOR]

In the context of climate change, large-scale net primary productivity (NPP) estimation and its impact feature variables are drawing more and more attention. Traditional process-based and empirical models are limited by their model structure and input variable design. The available field measurement data are limited by their small coverage. We propose to train an NPP calculation model using random forest and quantify the influence of multiple meteorological features on NPP. The calculated NPP correlates well with the MODIS product (correlation coefficient higher than 0.8). The importance rankings of multiple features are related to the local economy and development strategy of research areas. In developed areas, vegetation indexes are the most important, while in developing areas, land classification type influences NPP the most. The experiments suggest random forest is promising for estimating NPP and useful in analysing the impact features in terms of global change. [ABSTRACT FROM AUTHOR]

The low-resolution characteristic of passive microwave surface soil moisture (SSM) products greatly limits their application in many fields at regional or local scale. Aiming to overcome this limitation, a random forest (RF)-based downscaling approach was proposed in this study to disaggregate the Soil Moisture Active and Passive (SMAP) SSM product with the synergistic use of the Optical/Thermal infrared (TIR) observations from the Moderate-Resolution Imaging Spectro-radiometer (MODIS) onboard the Terra and Aqua satellites. The Iberian Peninsula was selected as the study area during the period from 2015 to 2016. First, the performance of the RF-based approach in building the SSM relationship model with surface variables (surface temperature, vegetation index, leaf area index, albedo, water index, solar factor, and elevation) was compared with that resulting from a widely used polynomial-based relationship model. Good agreement was achieved for the RF-based method with a correlation coefficient (R) above 0.95 and a mean root-mean-square deviation (RMSD) of 0.009 m 3 /m 3 . Next, four data combinations (AM + Terra, AM + Aqua, PM + Terra, and PM + Aqua) were generated according to the different overpass times of the SMAP and MODIS observations, and they were separately used to derive the spatially downscaled SSM with the proposed RF-based downscaling method. Validation was performed with the in situ measurements from the REMEDHUS network of the University of Salamanca in Spain. The results indicated that all combinations have similar good performances with an unbiased root-mean-square difference (ubRMSD) of 0.022 m 3 /m 3 , and the downscaled SSM at 1-km spatial resolution presented better accuracy while showing higher spatial heterogeneity and more detailed temporal pattern. Finally, the temporal changing pattern of the downscaled SSM was assessed with the precipitation time series from eight meteorological stations in the study area, and the rainfall effect on the variation of SSM was well tracked from its temporal changes. Overall, this study suggests that the proposed RF-based downscaling method is able to capture the variation of SSM well, and it should be helpful to improve the resolution of passive microwave soil moisture data and facilitate their applications at small scales. [ABSTRACT FROM AUTHOR]

Changes in the timing of plant phenological phases are important proxies in contemporary climate research. However, most of the commonly used traditional phenological observations do not give any coherent spatial information. While consistent spatial data can be obtained from airborne sensors and preprocessed gridded meteorological data, not many studies robustly benefit from these data sources. Therefore, the main aim of this study is to create and evaluate different statistical models for reconstructing, predicting, and improving quality of phenological phases monitoring with the use of satellite and meteorological products. A quality-controlled dataset of the 13 BBCH plant phenophases in Poland was collected for the period 2007-2014. For each phenophase, statistical models were built using the most commonly applied regression-based machine learning techniques, such as multiple linear regression, lasso, principal component regression, generalized boosted models, and random forest. The quality of the models was estimated using a k-fold cross-validation. The obtained results showed varying potential for coupling meteorological derived indices with remote sensing products in terms of phenological modeling; however, application of both data sources improves models’ accuracy from 0.6 to 4.6 day in terms of obtained RMSE. It is shown that a robust prediction of early phenological phases is mostly related to meteorological indices, whereas for autumn phenophases, there is a stronger information signal provided by satellite-derived vegetation metrics. Choosing a specific set of predictors and applying a robust preprocessing procedures is more important for final results than the selection of a particular statistical model. The average RMSE for the best models of all phenophases is 6.3, while the individual RMSE vary seasonally from 3.5 to 10 days. Models give reliable proxy for ground observations with RMSE below 5 days for early spring and late spring phenophases. For other phenophases, RMSE are higher and rise up to 9-10 days in the case of the earliest spring phenophases. [ABSTRACT FROM AUTHOR]

Forests in Honduras are endangered as a result of the relentless occurrence of wildfires during the dry season, and their frequency and area burned have been gradually increasing, a pattern attributable to the numerous ignition sources. For this reason, there is a substantial need to identify the major drivers of wildfires and map the regions where they are most likely to occur. In this study, we integrated the wildfire occurrences throughout the 2010–2015 period with a series of variables using the random forest algorithm. We included variables related to human activities such as the continuous distances to infrastructure and settlements. Other variables included are satellite observations that reflect the seasonal vegetation change, climatic conditions over the country, and topographical variables. The analysis of the explanatory variables revealed that the dry fuel conditions and low precipitation combined with the proximity to non-paved and paved roads were the major drivers of wildfires in the region. The estimated area with high and very high wildfire susceptibility was 15% of the country, located mainly in the central and eastern regions. The proposed national-scale wildfire susceptibility map can lead to enhanced preventive measures to minimize risk and the impacts caused by wildfires. [ABSTRACT FROM AUTHOR]

This paper aims to develop a global burned area (BA) algorithm for MODIS BRDF-corrected images based on the Random Forest (RF) classifier. Two RF models were generated, including: (1) all MODIS reflective bands; and (2) only the red (R) and near infrared (NIR) bands. Active fire information, vegetation indices and auxiliary variables were taken into account as well. Both RF models were trained using a statistically designed sample of 130 reference sites, which took into account the global diversity of fire conditions. For each site, fire perimeters were obtained from multitemporal pairs of Landsat TM/ETM+ images acquired in 2008. Those fire perimeters were used to extract burned and unburned areas to train the RF models. Using the standard MD43A4 resolution (500 × 500 m), the training dataset included 48,365 burned pixels and 6,293,205 unburned pixels. Different combinations of number of trees and number of parameters were tested. The final RF models included 600 trees and 5 attributes. The RF full model (considering all bands) provided a balanced accuracy of 0.94, while the RF RNIR model had 0.93. As a first assessment of these RF models, they were used to classify daily MCD43A4 images in three test sites for three consecutive years (2006-2008). The selected sites included different ecosystems: Australia (Tropical), Boreal (Canada) and Temperate (California), and extended coverage (totaling more than 2,500,000 km²). Results from both RF models for those sites were compared with national fire perimeters, as well as with two existing BA MODIS products; the MCD45 and MCD64. Considering all three years and three sites, commission error for the RF Full model was 0.16, with an omission error of 0.23. For the RF RNIR model, these errors were 0.19 and 0.21, respectively. The existing MODIS BA products had lower commission errors, but higher omission errors (0.09 and 0.33 for the MCD45 and 0.10 and 0.29 for the MCD64) than those obtained with the RF models, and therefore they showed less balanced accuracies. The RF models developed here should be applicable to other biomes and years, as they were trained with a global set of reference BA sites. [ABSTRACT FROM AUTHOR]

A five-year drought in California led to a significant increase in tree mortality in the Sierra Nevada forests from 2012 to 2016. Landscape level monitoring of forest health and tree dieback is critical for vegetation and disaster management strategies. We examined the capability of multispectral imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS) in detecting and explaining the impacts of the recent severe drought in Sierra Nevada forests. Remote sensing metrics were developed to represent baseline forest health conditions and drought stress using time series of MODIS vegetation indices (VIs) and a water index. We used Random Forest algorithms, trained with forest aerial detection surveys data, to detect tree mortality based on the remote sensing metrics and topographical variables. Map estimates of tree mortality demonstrated that our two-stage Random Forest models were capable of detecting the spatial patterns and severity of tree mortality, with an overall producer's accuracy of 96.3% for the classification Random Forest (CRF) and a RMSE of 7.19 dead trees per acre for the regression Random Forest (RRF). The overall omission errors of the CRF ranged from 19% for the severe mortality class to 27% for the low mortality class. Interpretations of the models revealed that forests with higher productivity preceding the onset of drought were more vulnerable to drought stress and, consequently, more likely to experience tree mortality. This method highlights the importance of incorporating baseline forest health data and measurements of drought stress in understanding forest response to severe drought. [ABSTRACT FROM AUTHOR]

Droughts, typically caused by the deficiencies of precipitation and soil moisture, affect water resources and agriculture. As soil moisture is of key importance in understanding the interaction between the atmosphere and Earth’s surface, it can be used to monitor droughts. In this study, a High resolution Soil Moisture Drought Index (HSMDI) was proposed and evaluated for meteorological, agricultural, and hydrological droughts. HSMDI was developed using the 1 km downscaled soil moisture data produced from the Advanced Microwave Scanning Radiometer on the Earth Observing System (AMSR-E) from 2003 to 2011 (March to November) over the Korean peninsula. Seven products from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Tropical Rainfall Measuring Mission (TRMM) satellite sensors were used to downscale AMSR-E soil moisture based on random forest machine learning. The downscaled 1 km soil moisture was correlated well with both in situ and AMSR-E soil moisture with the mean coefficient of determination ( R 2 ) of 0.29 and 0.59, respectively. The Standardized Precipitation Index (SPI) with time scales from 1 to 12 months, crop yields (for sesame, highland radish, and highland napa cabbage) and streamflow data were used to validate HSMDI for various types of droughts. The results showed that HSMDI depicted meteorological drought well, especially during the dry season, with a similar pattern with the 3-month SPI. However, the performance fluctuated a bit during the wet season possibly due to the limited availability of optical sensor data and heterogeneous land covers around the stations. HSMDI also showed high correlation with crop yield data, in particular the highland radish and napa cabbage cultivated in non-irrigated regions with a mean R 2 of 0.77. However, HSMDI did not monitor streamflow well for hydrological drought presenting a various range of correlations with streamflow data (from 0.03 to 0.83). [ABSTRACT FROM AUTHOR]

The Asian Migratory locust ( Locusta migratoria migratoria L.) is a pest that continuously threatens crops in the Amudarya River delta near the Aral Sea in Uzbekistan, Central Asia. Its development coincides with the growing period of its main food plant, a tall reed grass (Phragmites australis), which represents the predominant vegetation in the delta and which cover vast areas of the former Aral Sea, which is desiccating since the 1960s. Current locust survey methods and control practices would tremendously benefit from accurate and timely spatially explicit information on the potential locust habitat distribution. To that aim, satellite observation from the MODIS Terra/Aqua satellites and in-situ observations were combined to monitor potential locust habitats according to their corresponding risk of infestations along the growing season. A Random Forest (RF) algorithm was applied for classifying time series of MODIS enhanced vegetation index (EVI) from 2003 to 2014 at an 8-day interval. Based on an independent ground truth data set, classification accuracies of reeds posing a medium or high risk of locust infestation exceeded 89% on average. For the 12-year period covered in this study, an average of 7504 km 2 (28% of the observed area) was flagged as potential locust habitat and 5% represents a permanent high risk of locust infestation. Results are instrumental for predicting potential locust outbreaks and developing well-targeted management plans. The method offers positive perspectives for locust management and treatment of infested sites because it is able to deliver risk maps in near real time, with an accuracy of 80% in April-May which coincides with both locust hatching and the first control surveys. Such maps could help in rapid decision-making regarding control interventions against the initial locust congregations, and thus the efficiency of survey teams and the chemical treatments could be increased, thus potentially reducing environmental pollution while avoiding areas where treatments are most likely to cause environmental degradation. [ABSTRACT FROM AUTHOR]

• 30 m resolution annual soy yield maps produced over Brazil 2001–2020. • Map RMSE 418 kg/ha or 15% relative to long-term average yield. • Per-pixel uncertainty estimated for annual yield maps. • Trained models can be used for operational prediction in future years. Long-term spatially explicit information on crop yield is essential for understanding food security in a changing climate. Here we present a study that combines twenty-years of Landsat and MODIS data, climate and weather records, municipality-level crop yield statistics, random forests and linear regression models for mapping crop yield in a multi-temporal, multi-scale modeling framework. The study was conducted for soybean in Brazil. Using a recently developed 30 m resolution, annual (2001–2019) soybean classification map product, we aggregated multi-temporal phenological metrics derived from Landsat and MODIS data over soybean pixels to the municipality scale. We combined phenological metrics with topographic features, long-term climate data, in-season weather data and soil variables as inputs to machine learning models. We trained a multi-year random forests model using yield statistics as reference and subsequently applied linear regression to adjust the biases in the direct output of the random forests model. This model combination achieved the best performance with a root-mean-square-error (RMSE) of 344 kg/ha (12% relative to long-term mean yield) and an r2 of 0.69, on the basis of 20% withheld test data. The RMSE of the leave-one-year-out model assessment ranged from 259 kg/ha to 816 kg/ha. To eliminate the artifacts caused by the coarse-resolution climate and weather data, we developed multiple models with different categories of input variables. Employing the per-pixel uncertainty estimates of different models, the final soybean yield maps were produced through per-pixel model composition. We applied the models trained on 2001–2019 data to 2020 data and produced a soybean yield map for 2020, demonstrating the predictive capability of trained machine learning models for operational yield mapping in future years. Our research showed that combining satellite, climate and weather data and machine learning could effectively map crop yield at high resolution, providing critical information to understand yield growth, anomaly and food security. [ABSTRACT FROM AUTHOR]

With the gradual declining streamflow gauging stations in many world-rivers, emphasis is given nowadays to develop remote sensing (RS)-based approaches as the next-generation hydrometry for estimating riverine ecological flow regimes (EFR). For constructing EFR based on daily-streamflow data in scantily-gauged reaches, use of RS techniques in narrow flow-width tropical rain-fed rivers is constrained with the non-availability of finer spatial satellite data at daily scale. To address these limitations, this study proposes a novel framework that integrates the enhanced spatiotemporal adaptive reflectance fusion (FUS) of the 250 m × 1-day resolution Aqua-MODIS and 30 m × 1-day resolution Landsat satellite-based remote sensing images in the near-infrared region with the machine learning algorithms. These developed frameworks are named as Artificial Neural Network-based ANNFUS, Random Forest Regression-based RFRFUS, and Support Vector Regression-based SVRFUS models, which were tested for daily-scale streamflow estimation in a typical Brahmani River Basin, India. The results reveal that by addressing the linear and nonlinear dynamism between the streamflow and satellite signals, all the developed models could simulate the streamflow very well with the Nash-Sutcliffe efficiency>0.8, Kling-Gupta efficiency>0.8, relative root mean square error (rRMSE) of 0.051–0.12, and normalized RMSE of 0.23–0.36. However, for reproducing the high, median, and low streamflow regimes, the SVRFUS model was found to be the best with the NSE>0.85 and KGE>0.8. Conclusively, the proposed approach is found to have the potential to be replicated in other world-river basins to estimate ecological flow regimes at defunct gauging stations facilitating the basin-scale aquatic environmental management. [Display omitted] • To estimate daily streamflow, advocated remote sensing-based machine learning tools. • The models are ANN, Random Forest Regression and Support Vector Regression. • For river/near-river environments, MODIS and Landsat pixels in NIR-band are fused. • Tested the proposed techniques in a narrow-width river constrained with mixels. • For streamflow regimes, Support Vector Regression models are the best. [ABSTRACT FROM AUTHOR]

Environmental monitoring with satellite data is facilitated by frequent observations at a fine spatial scale. As land surface temperature (LST) is one environmental key variable, we implemented a random forest (RF) regression approach to increase the spatial resolution of LST maps from ~ 1 km, routinely available in daily repetition from the Moderate Resolution Imaging Spectroradiometer (MODIS), to ~ 250 m. LST was downscaled based on its relationship to topographic variables derived from digital elevation data of the Shuttle Radar Topography Mission (SRTM), land cover data (MODIS product MCD12Q1), and surface reflectances in the visible red and near-infrared, which both are provided with the MODIS/Terra daily product MOD09GQ at ~ 250 m resolution. The approach was tested for a complex landscape in the Eastern Mediterranean, the Jordan River Region, with LST fields from aggregated Landsat-7 ETM + and MODIS (MODIS/Terra LST product MOD11A1) data; as reference at the finer scale, we used Landsat-7 derived LST data. For the ideal-case scenario with both degraded and reference values from the same sensor (Landsat-7 ETM +), root mean square errors (RMSE) of downscaled LSTs ranged from 1.02 K to 1.43 K for six different acquisition dates. When compared to the widely-adopted and in parallel applied TsHARP sharpening method, that is based on the relationship between NDVI and LST, downscaling accuracy with RF improved up to 19%. Applying the RF approach to MODIS LST products yielded RMSEs from 1.41 K to 1.92 K, whereas the TsHARP method and also a uniform disaggregation by resampling provided only slightly worse results. For the real MODIS LST product, downscaling with RF was affected by lower thermal contrasts in the image data that hindered an adequate training to reproduce temperature variations at the finer scale of ~ 250 m. In this context, we assume the LST product of the Visible Infrared Imaging Radiometer Suite (VIIRS) instrument (as a successor of MODIS onboard the SUOMI NPP platform) to be a better candidate for downscaling, as it provides a spatial resolution of ~ 750 m. One advantage of the RF approach is that predictor datasets can easily be adapted to data availability. In an extended RF approach with all Landsat spectral bands, downscaling results for formerly aggregated Landsat data improved distinctly and now ranged from 0.98 K to 1.33 K. This approach is also promising for the downscaling of real MODIS or VIIRS LST data as it may be combined with already available reflective data fusion models that are able to blend Landsat data and sensor data with a coarse spatial resolution (given spectral bandwidths corresponding to Landsat) to generate temporally dense synthetic Landsat time series. [ABSTRACT FROM AUTHOR]

The Yellow River Basin is one of China's most densely-populated, fastest growing and most dynamic regions, with abundant natural resources and intense agricultural production. Major land policies have recently resulted in remarkable landscape modifications throughout the basin. The availability of precise regional land cover change information is crucial to better understand the prevailing dynamics and underlying factors influencing the current processes in such a complex system and can additionally serve as a valuable component for modeling and decision making. Such comprehensive and detailed information is lacking for the Yellow River Basin so far. In this study, we derived land cover characteristics and dynamics from the complete last decade based on optical high-temporal MODIS Normalized Differenced Vegetation Index (NDVI) time series for the whole Yellow River Basin. After filtering and smoothing for noise reduction with the use of the adaptive Savitzky-Golay filter, the processed time series was used to derive a large variety of phenological and annual metrics. The final classifications for the basin (2003 and 2013) were based on a random forest classifier, trained by reference samples from very high-resolution imagery. The accuracy assessment for all 18 thematic classes, which was based on a 30% reference data split, yielded an overall accuracy of 87% and 84% for 2003 and 2013, respectively. Major land cover and land use changes during the last decade have occurred on the Loess Plateau, where land and conservation reforms triggered large-scale recovery of grassland and shrubland habitat that had been previously covered by agriculture or sparse vegetation. Agricultural encroachment and urban area expansion are other processes influencing the dynamics in the basin. The necessity for regionally-adapted land cover maps becomes obvious when our land cover products are compared to existing global products, where thematic accuracy remains low, particularly in a heterogeneous landscape, such as the Yellow River Basin. The basin-wide novel land cover and land use products of the Yellow River Basin hold a large potential for climate, hydrology and biodiversity modelers, as well as river basin and regional governmental authorities and will be shared upon request. [ABSTRACT FROM AUTHOR]

Regular monitoring of expanding impervious surfaces areas (ISAs) in urban areas is highly desirable. MODIS data can meet this demand in terms of frequent observations but are lacking in spatial detail, leading to the mixed land cover problem when per-pixel classifications are applied. To overcome this issue, this research develops and applies a spatio-temporal sub-pixel model to estimate ISAs on an annual basis during 2001-2013 in the Jakarta Metropolitan Area, Indonesia. A Random Forest (RF) regression inferred the ISA proportion from annual 23 values of MODIS MOD13Q1 EVI and reference data in which such proportion was visually allocated from very high-resolution images in Google Earth over time at randomly selected locations. Annual maps of ISA proportion were generated and showed an average increase of 30.65 km²/year over 13 years. For comparison, a series of RF per-pixel classifications were also developed from the same reference data using a Boolean class constructed from different thresholds of ISA proportion. Results from per-pixel models varied when such thresholds change, suggesting difficulty of estimation of actual ISAs. This research demonstrated the advantages of spatio-temporal sub-pixel analysis for annual ISAs mapping and addresses the problem associated with definitions of thresholds in per-pixel approaches. [ABSTRACT FROM AUTHOR]

Agricultural irrigation, as an important practice to protect crops from drought and promote grain yield, has a long history in China. A timely and precise dataset about the extent and dynamics of irrigated areas is necessary for water allocation and agricultural management but is scarce in China. Here we developed annual irrigated cropland maps across China (IrriMap_CN) at 500-m resolution from 2000 to 2019, using MODIS data, machine-learning method, and Google Earth Engine platform. First, we generated annual nationwide training samples by strictly screening the existing irrigation maps downscaled from the statistical data. Second, we implemented locally adaptive random forest classifiers in 511 nominal 1° × 1° grid cells across China with MODIS vegetation indices, climatic factors, and topography variables. Third, we conducted nationwide pixel-wise validation of the IrriMap_CN using independent samples. The validation results based on more than 3000 ground truth points revealed that IrriMap_CN had high accuracies ranging from 77.2% to 85.9%. The time series of IrriMap_CN detected substantial expansion of irrigated areas in Xinjiang and Heilongjiang (more than 50% in total) and pronounced decreases in Sichuan, Jiangsu, and Hebei. The analyses of irrigation frequency, start time, and end time implied that North China Plain was the most intensive irrigated area; but the irrigation area showed a decreasing trend since 2000, consistent with the reduced agricultural water consumption. The annual irrigation datasets allow us to understand the spatiotemporal dynamics of irrigated croplands in China and are expected to contribute to the improvement of earth system models and facilitate sustainable agricultural water management. • Annual irrigation maps (IrriMap_CN) were generated for China in 2000–2019. • Nationwide training samples were extracted from existing irrigation maps. • IrriMap_CN highlights the declining irrigation area in North China Plain. • Cropland reclamation/occupation and water supply are key to irrigation area changes. [ABSTRACT FROM AUTHOR]

Mapping urban areas at global and regional scales is an urgent and crucial task for detecting urbanization and human activities throughout the world and is useful for discerning the influence of urban expansion upon the ecosystem and the surrounding environment. DMSP-OLS stable nighttime lights have provided an effective way to monitor human activities on a global scale. Threshold-based algorithms have been widely used for extracting urban areas and estimating urban expansion, but the accuracy can decrease because of the empirical and subjective selection of threshold values. This paper proposes an approach for extracting urban areas with the integration of DMSP-OLS stable nighttime lights and MODIS data utilizing training sample datasets selected from DMSP-OLS and MODIS NDVI based on several simple strategies. Four classification algorithms were implemented for comparison: the classification and regression tree (CART), k-nearest-neighbors (k-NN), support vector machine (SVM), and random forests (RF). A case study was carried out on the eastern part of China, covering 99 cities and 1,027,700 km2. The classification results were validated using an independent land cover dataset, and then compared with an existing contextual classification method. The results showed that the new method can achieve results with comparable accuracies, and is easier to implement and less sensitive to the initial thresholds than the contextual method. Among the four classifiers implemented, RF achieved the most stable results and the highest average Kappa. Meanwhile CART produced highly overestimated results compared to the other three classifiers. Although k-NN and SVM tended to produce similar accuracy, less-bright areas around the urban cores seemed to be ignored when using SVM, which led to the underestimation of urban areas. Furthermore, quantity assessment showed that the results produced by k-NN, SVM, and RFs exhibited better agreement in larger cities and low consistency in small cities. [ABSTRACT FROM AUTHOR]

Dust storms are important phenomena over large regions of the arid and semi-arid areas of the Middle East. Due to the influences of dust aerosols on climate and human daily activities, dust detection plays a crucial role in environmental and climatic studies. Detection of dust storms is critical to accurately understand dust, their properties and distribution. Currently, remotely sensed data such as MODIS (Moderate Resolution Imaging Spectroradiometer) with appropriate temporal and spectral resolutions have been widely used for this purpose. This paper investigates the capability of two physical-based methods, and random forests (RF) classifier, for the first time, to detect dust storms using MODIS imagery. Since the physical-based approaches are empirical, they suffer from certain drawbacks such as high variability of thresholds depending on the underlying surface. Therefore, classification-based approaches could be deployed as an alternative. In this paper, the most relevant bands are chosen based on the physical effects of the major classes, particularly dust, cloud and snow, on both emissive infrared and reflective bands. In order to verify the capability of the methods, OMAERUV AAOD (aerosol absorption optical depth) product from OMI (Ozone Monitoring Instrument) sensor is exploited. In addition, some small regions are selected manually to be considered as ground truth for measuring the probability of false detection (POFD) and probability of missing detection (POMD). The dust class generated by RF is consistent qualitatively with the location and extent of dust observed in OMERAUV and MODIS true colour images. Quantitatively, the dust classes generated for eight dust outbreaks in the Middle East are found to be accurate from 7% and 6% of POFD and POMD respectively. Moreover, results demonstrate the sound capability of RF in classifying dust plumes over both water and land simultaneously. The performance of the physical-based approaches is found weaker than RF due to empirical thresholds that are not always true. [ABSTRACT FROM AUTHOR]

Currently, accurate information on crop area coverage is vital for food security and industry, and there is strong demand for timely crop mapping. In this study, we used MODIS time series data to investigate the effect of the time series length on crop mapping. Eight time series with different lengths (ranging from one month to eight months) were tested. For each time series, we first used the Random Forest (RF) algorithm to calculate the importance score for all features (including multi-spectral data, Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and phenological metrics). Subsequently, an extension of the Jeffries-Matusita (JM) distance was used to measure class separability for each time series. Finally, the RF algorithm was used to classify crop types, and the classification accuracy and certainty were used to analyze the influence of the time series length and the number of features on classification performance; the features were added one by one based on their importance scores. Results indicated that when the time series was longer than five months, the top ten features remained stable. These features were mainly in July and August. In addition, the NDVI features contributed the majority of the most significant features for crop mapping. The NDWI and data from multi-spectral bands also contributed to improving crop mapping. On the other hand, separability, classification accuracy, and certainty increased with the number of features used and the time series length, although these values quickly reached saturation. Five months was the optimal time series length, as longer time series provided no further improvement in the classification performance. This result shows that relatively short time series have the potential to identify crops accurately, which allows for early crop mapping over large areas. [ABSTRACT FROM AUTHOR]

The analysis and classification of land cover is one of the principal applications in terrestrial remote sensing. Due to the seasonal variability of different vegetation types and land surface characteristics, the ability to discriminate land cover types changes over time. Multi-temporal classification can help to improve the classification accuracies, but different constraints, such as financial restrictions or atmospheric conditions, may impede their application. The optimisation of image acquisition timing and frequencies can help to increase the effectiveness of the classification process. For this purpose, the Feature Importance (FI) measure of the state-of-the art machine learning method Random Forest was used to determine the optimal image acquisition periods for a general (Grassland, Forest, Water, Settlement, Peatland) and Grassland specific (Improved Grassland, Semi-Improved Grassland) land cover classification in central Ireland based on a 9-year time-series of MODIS Terra 16 day composite data (MOD13Q1). Feature Importances for each acquisition period of the Enhanced Vegetation Index (EVI) and Normalised Difference Vegetation Index (NDVI) were calculated for both classification scenarios. In the general land cover classification, the months December and January showed the highest, and July and August the lowest separability for both VIs over the entire nine-year period. This temporal separability was reflected in the classification accuracies, where the optimal choice of image dates outperformed the worst image date by 13% using NDVI and 5% using EVI on a mono-temporal analysis. With the addition of the next best image periods to the data input the classification accuracies converged quickly to their limit at around 8–10 images. The binary classification schemes, using two classes only, showed a stronger seasonal dependency with a higher intra-annual, but lower inter-annual variation. Nonetheless anomalous weather conditions, such as the cold winter of 2009/2010 can alter the temporal separability pattern significantly. Due to the extensive use of the NDVI for land cover discrimination, the findings of this study should be transferrable to data from other optical sensors with a higher spatial resolution. However, the high impact of outliers from the general climatic pattern highlights the limitation of spatial transferability to locations with different climatic and land cover conditions. The use of high-temporal, moderate resolution data such as MODIS in conjunction with machine-learning techniques proved to be a good base for the prediction of image acquisition timing for optimal land cover classification results. [ABSTRACT FROM AUTHOR]

Anthropogenic climate and land use change is causing rapid shifts in the distribution and composition of habitats with profound impacts on ecosystem biodiversity. The sustainable management of ecosystems requires monitoring programmes capable of detecting shifts in habitat distribution and composition at large spatial scales. Remote sensing observations facilitate such efforts as they enable cost-efficient modelling approaches that utilize publicly available datasets and can assess the status of habitats over extended periods of time. In this study, we introduce a modelling framework for habitat monitoring in Germany using readily available MODIS surface reflectance data. We developed supervised classification models that allocate (semi-)natural areas to one of 18 classes based on their similarity to Natura 2000 habitat types. Three machine learning classifiers, i.e., Support Vector Machines (SVM), Random Forests (RF), and C5.0, and an ensemble approach were employed to predict habitat type using spectral signatures from MODIS in the visible-to-near-infrared and short-wave infrared. The models were trained on homogenous Special Areas of Conservation that are predominantly covered by a single habitat type with reference data from 2013, 2014, and 2016 and tested against ground truth data from 2010 and 2019 for independent model validation. Individually, the SVM and RF methods achieved better overall classification accuracies (SVM: 0.72–0.93%, RF: 0.72–0.94%) than the C5.0 algorithm (0.66–0.93%), while the ensemble classifier developed from the individual models gave the best performance with overall accuracies of 94.23% for 2010 and 80.34% for 2019 and also allowed a robust detection of non-classifiable pixels. We detected strong variability in the cover of individual habitat types, which were reduced when aggregated based on their similarity. Our methodology is capable to provide quantitative information on the spatial distribution of habitats, differentiate between disturbance events and gradual shifts in ecosystem composition, and could successfully allocate natural areas to Natura 2000 habitat types. [ABSTRACT FROM AUTHOR]

The increasing wildfire occurrence due to global climate changes urged the improvement of present wildfire growth prediction and evaluation methods. This study aimed to propose novel solutions to their two primary limitations, including the lack of robust fuel classification method and the low spatial resolution of wildfire growth accuracy assessment while ensuring wide applicability using open data satellite missions and software. The first objective was to create a robust two-step fuel model classification method consisted of the supervised machine learning classification of generalized land cover classes in the 1st level and their individual unsupervised classification to vegetation subtypes in the 2nd level. The second objective was creating a wildfire prediction accuracy assessment method using MODIS 250 m images, which overcome the limitations of low spatial resolution while preserving sub-daily temporal resolution. The wildfire on the Korčula island in Croatia was analyzed in the study, being specific for its long duration from 18 to 24 July 2015. The wildfire ignition occurred in the isolated area, which prolonged the response time from emergency agencies. Random Forest (RF) with input Landsat 8 spectral bands and indices resulted in the highest classification accuracy in the 1st classification level with an overall agreement of 83.6%. The vegetation subclasses from the 2nd classification level were matched to the 13 standard fuel models for the input in FARSITE software. The predicted wildfire evaluation showed the highest mean accuracy of 0.906 for the first two days, which decreased to 0.722 in the latter stages of the active wildfire caused by overprediction. The proposed two-step fuel model classification presented a cost-efficient solution to the fuel map creation in any part of the world, with a disadvantage of no in-situ ground truth identification and accuracy assessment for 2nd classification level. The evaluation of wildfire growth prediction with 250 m images enabled high spatial and temporal resolution of the assessment, while its limitations of wildfire overprediction and the negative effects of wildfire smoke in MODIS images should be addressed in future research. [Display omitted] • Wildfire growth was predicted and evaluated using Landsat and MODIS images. • The fuel model was classified using a robust two-step classification method. • Wildfire perimeters were evaluated using 250 m daily NDVI time series. • FARSITE prediction mean accuracy was 0.906 in the first two days of active wildfire. • Proposed cost-efficient methods support the decision-making in web services. [ABSTRACT FROM AUTHOR]

The surface all-wave net radiation ( R n ) plays an important role in the energy and water cycles, and most studies of R n estimations have been conducted using satellite data. As one of the most commonly used satellite data sets, Moderate Resolution Imaging Spectroradiometer (MODIS) data have not been widely used for radiation calculations at mid-low latitudes because of its very low revisit frequency. To improve the daily R n estimation at mid-low latitudes with MODIS data, four models, including three models built with random forest (RF) and different temporal expansion models and one model built with the look-up-table (LUT) method, are used based on comprehensive in situ radiation measurements collected from 340 globally distributed sites, MODIS top-of-atmosphere (TOA) data, and the fifth generation of European Centre for Medium-Range Weather Forecasts Reanalysis 5 (ERA5) data from 2000 to 2017. After validation against the in situ measurements, it was found that the RF model based on the constraint of the daily R n from ERA5 (an RF-based model with ERA5) performed the best among the four proposed models, with an overall validated root-mean-square error (RMSE) of 21.83 Wm−2, R2 of 0.89, and a bias of 0.2 Wm−2. It also had the best accuracy compared to four existing products (Global LAnd Surface Satellite Data (GLASS), Clouds and the Earth's Radiant Energy System Edition 4A (CERES4A), ERA5, and FLUXCOM_RS) across various land cover types and different elevation zones. Further analyses illustrated the effectiveness of the model by introducing the daily R n from ERA5 into a "black box" RF-based model for R n estimation at the daily scale, which is used as a physical constraint when the available satellite observations are too limited to provide sufficient information (i.e., when the overpass time is less than twice per day) or the sky is overcast. Overall, the newly-proposed RF-based model with ERA5 in this study shows satisfactory performance and has strong potential to be used for long-term accurate daily R n global mapping at finer spatial resolutions (e.g., 1 km) at mid-low latitudes. [ABSTRACT FROM AUTHOR]