Search

There are challenges for cloud detection over bright surfaces, especially for Sentinel-2 image without thermal band. In this paper, we developed a new cloud detection algorithm for Sentinel-2 data designed to tackle the challenges of cloud detection on bright targets. Our algorithm (called RF-NDTeI-SNIC) was built through the integration of Random Forest (RF) model, Normalized Difference Temporal Index (NDTeI) cloud index and Simple Non-Iterative Clustering (SNIC) image segmentation algorithm. First, RF and NDTeI were used to derive the initial and second stage's cloud detection results, respectively. Then, the RF- and NDTeI-based results were fused to remove the bright surfaces and land cover changes. Finally, the fusion results were further refined using SNIC for morphological processing. The proposed RF-NDTeI-SNIC algorithm was evaluated using nine challenging cloud-covered Sentinel-2 scenes with substantial bright surfaces. Results indicated that the average overall accuracy of our algorithm was 95.01%, with average commission rate of 10.19% for cloud. In addition, five commonly used Sentinel-2 cloud mask algorithms including s2cloudless, QA, Fmask, CDI and Tmask were selected for comparative analysis. Results suggested that our algorithm outperformed others with overall accuracies of 89.60–94.17% and cloud commission rate of 9.46–37.50%, and have significant advantages in terms of bright surfaces. In summary, the RF-NDTeI-SNIC algorithm we developed was capable of yielding accurate Sentinel-2 cloud masks, and the novel NDTeI cloud index we proposed gave a new and effective approach for improving the separability of clouds and bright surfaces. [ABSTRACT FROM AUTHOR]

The Global Ecosystem Dynamics Investigation (GEDI) is a high-resolution lidar instrument, facilitating landscape monitoring through forest structure and biomass assessments. Previous research has mainly focused on estimating forest attributes such as height and biomass from GEDI data. In this comprehensive study, we empirically investigate how Eucalyptus plantation characteristics and environmental factors collectively influence GEDI waveform metrics. Using a diverse dataset that includes measurements of canopy height, planting density, foliage characteristics, species composition, understorey presence and environmental conditions such as climate and soil properties, we explore the complex relationships between these attributes and GEDI waveform metrics using a two-step approach that includes assessing the linear relationship between each in-situ parameter and GEDI metrics, and also random forest modelling. Key findings include: 1) While canopy height (Ht) plays a significant role in shaping the GEDI waveform, forest wood volume (Volume), which incorporates height and diameter at breast height (DBH) of the trees and also tree stem density (Ntree/ha), exerts an even greater influence on GEDI metrics, in particular on the waveform extent (WE) and relative heights (RH); 2) The influence of factors such as foliage biomass and wood volume on GEDI waveform varies vertically from the forest floor up to forest top, and the model was more precise to predict the top part of the waveform; 3) No single attribute can solely account for the observed variations in the waveform characteristics, but multiple interactions between different forest and environmental features contribute to the complex patterns in the received waveforms. This research contributes to a deeper understanding of the complex relationships between forest plantation characteristics and GEDI waveform metrics and highlights the need to consider a wide range of forest attributes in the analyse of GEDI waveform to produce more accurate canopy height or volume estimates. [ABSTRACT FROM AUTHOR]

Spectral diversity (SD) in reflectance can be used to estimate plant taxonomic diversity (TD) according to the Spectral Variation Hypothesis (SVH). However, contrasting relationships between SD and TD have been reported by different studies. Indeed, multiple factors may affect SD, including spatial and spectral scales, vegetation characteristics and the adopted SD computational method. Here, we tested the SVH over 171 plots within a large and heterogeneous forest area in North-Eastern Italy using Sentinel-2 data, aiming at identifying possible factors affecting the strength and direction of SD-TD relationship. SD was determined using 'biodivMapR' (BD) and 'rasterdiv' (RD) R packages and 38 possible combinations of SD indices, at both α (within a community) and β (among communities) levels, and computational parameters accounting for spatial and spectral scales. Information on vegetation structure was either retrieved from ground-based or LiDAR data. A Random Forest approach was used to disentangle the relationships between SD, TD and vegetation structure, and to identify the best combination of SD computational parameters. At the α-level, we found negative relationship between TD and RD SD indices, which was mainly driven by the presence of gaps within the forest canopy. As regards BD, we found that this algorithm reduced background contribution on SD and was able to differentiate major forest types (broadleaves vs conifers), but derived α-SD indices were marginally correlated with α-TD. At the β-level, we observed a statistically significant positive correlation between BD SD indices and TD (maximum r = 0.24). Finally, we found stronger correlations and R2 when SD indices were calculated using smaller computation windows and over a larger pixels extraction area. Our findings suggest that vegetation cover and structure play a major role, with respect to inter-species spectral differences, in determining α-SD, and that SD might better capture differences in species composition at the landscape-level rather than the richness of individual communities. [ABSTRACT FROM AUTHOR]

As valued members of forest ecosystems, fast-growing tree species are of great importance for protecting natural forests, fulfilling the demand for raw wood material, and preserving biodiversity and ecosystem services. In this regard, multi-temporal monitoring and evaluation of poplar planted areas are essential for decision makers and planners to keep the inventory maps of planted areas up-to-date and track the temporal changes accurately. Using a spectral library, available spectral knowledge can be incorporated into inventory mapping and image classification processes. In this study, in situ spectral measurements of five Populus (four clones and one species) planted sites in Türkiye were collected at three phenological stages (early, mid, and late vegetation) and a novel spectral library was created after completing strict processing stages. Analysis of the spectral reflectances of the Populus species revealed that their spectral properties in the visible region are similar, whereas the discrimination in the near-infrared and shortwave infrared (SWIR) regions was noticeably higher. Two types of applications were conducted to effectively use the spectral library. First, the spectral library was used to validate the poplar class labels of the sites visited during the field campaign. The effectiveness of Worldview-3 imagery with eight SWIR bands was evaluated by pairwise comparison of the spectral curves of the sample fields with those of the spectral library. Second, the spectral library facilitated the extraction of poplar samples from Sentinel-2 imagery in an unvisited region, and a supervised classification process was conducted using the random forest algorithm. The results revealed that the Populus species (Samsun and I-214) could be delineated with high accuracy (>70%) using 10-band Sentinel-2 imagery, and object-based classification outperformed pixel-based classification by approximately 5% for the poplar types. The developed spectral library stands as a valuable asset for studies focused on identifying and classifying Populus species using high-resolution imagery. [ABSTRACT FROM AUTHOR]

Hyperspectral imaging effectively distinguishes land cover classes with discriminative spectral-spatial characteristics. However, hyperspectral image (HSI) classification encounters three significant challenges: the curse of dimensionality, the scarcity of training samples, and the necessity to incorporate spatial information within the classification process. The large number of redundant and irrelevant spectral bands in HSI can negatively impact the classification performance. Furthermore, the annotation process for training samples is expensive and time-consuming. To address these challenges, this work presents a novel semi-supervised approach for hyperspectral image dimensionality reduction using the Mayfly Algorithm (MA). The proposed method aims to preserve the most informative spectral bands while leveraging both labelled and unlabelled samples in the process. The MA algorithm effectively balances exploration and exploitation strategies, enhancing population diversity and mitigating challenges such as premature convergence and local optima. Additionally, the inclusion of structural variation and thresholding (SVT) in the Bitonic filter enables the incorporation of spatial features during the classification phase. This improves the separability of pixels belonging to different objects and leads to significant improvements in classification performance. Finally, two machine learning classifiers, namely Random Forest (RF) and Support Vector Machine (SVM), are applied at the pixel level for hyperspectral image classification. Extensive experimental results on three real hyperspectral datasets demonstrate the superiority of the proposed approach compared to state-of-the-art algorithms like Harmony Search (HS), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and more with an average improvement of 3.2%. [ABSTRACT FROM AUTHOR]

Detecting building rooftops with very high-resolution (VHR) images is an important issue in many fields, including disaster management, urban planning, and climate change research. Buildings with varying geometrical features are challenging to detect accurately from VHR image due to complicated image scenes containing spectrally similar objects, illumination, occlusions, viewing angles, and shadows. This study aims to detect building rooftops with high accuracy using a new framework that includes VHR image, visible band difference vegetation index, digital surface and elevation models, the terrain ruggedness and the topographic position index. Five distinct feature sets were generated in order of importance by exposing the ten related stacking features to a feature selection procedure using the maximum relevance minimum redundancy method. Then, Auto-Encoder, k-NN, decision tree, RUSBoost, and random forest machine learning algorithms were utilized for binary classification. Random forest yielded the highest accuracy (97.2% F-score, 98.72% accuracy) when all features (F10) were used, while decision tree was the least successful (59.16% F-score, 83.56% accuracy) for RGB feature set (FRGB). It was revealed that classification of F10 with random forest increased F-score by about 23% compared to classification with FRGB. Additionally, McNemar's tests showed no statistically significant difference between random forest vs k-NN and decision tree vs RUSBoost. [ABSTRACT FROM AUTHOR]

Since forest fire is one of the most dangerous natural disasters and presents serious threats to the local ecological environment, economic growth, and public safety, it is essential to carry out accurate and real-time forest fire monitoring. In this study, a forest fire monitoring method that combines a threshold-based algorithm and random forest (RF) model using Himawari-8 data is proposed. The threshold-based algorithm employs the solar zenith angle to adaptively determine the potential fire point judgement threshold to extract possible fire points. The RF model constructed with spectral features and spatio-temporal information is subsequently utilized to eliminate pseudo-fire points from the results of the threshold-based algorithm. To eliminate fire points in non-forest areas, post-processing is performed using land cover data. Five fire occurrence moments in the research area are selected to verify the identification accuracy. The results reveal that the overall accuracy and the overall comprehensive evaluation values are 97.36% and 0.913, respectively, which demonstrates that the proposed method is capable of accurately identifying forest fire points and providing an effective means for forest fire monitoring. [ABSTRACT FROM AUTHOR]

Olives are a crucial economic crop in Mediterranean countries. Detailed spatial information on the distribution and condition of crops at regional and national scales is essential to ensure the continuity of crop quality and yield efficiency. However, most earlier studies on olive tree mapping focused mainly on small parcels using single-sensor, very high resolution (VHR) data, which is time-consuming, expensive and cannot feasibly be scaled up to a larger area. Therefore, we evaluated the performance of Sentinel-1 and Sentinel-2 data fusion for the regional mapping of olive trees for the first time, using the Izmir Province of Türkiye, an ancient olive-growing region, as a case study. Three different monthly composite images reflecting the different phenological stages of olive trees were selected to separate olive trees from other land cover types. Seven land-cover classes, including olives, were mapped separately using a random forest classifier for each year between 2017 and 2021. The results were assessed using the k-fold cross-validation method, and the final olive tree map of Izmir was produced by combining the olive tree distribution over two consecutive years. District-level areas covered by olive trees were calculated and validated using official statistics from the Turkish Statistical Institute (TUIK). The K-fold cross-validation accuracy varied from 94% to 95% between 2017 and 2021, and the final olive map achieved 98% overall accuracy with 93% producer accuracy for the olive class. The district-level olive area was strongly related to the TUIK statistics (R2 = 0.60, NRMSE = 0.64). This study used Sentinel data and Google Earth Engine (GEE) to produce a regional-scale olive distribution map that can be scaled up to the entire country and replicated elsewhere. This map can, therefore, be used as a foundation for other scientific studies on olive trees, particularly for the development of effective management practices. [ABSTRACT FROM AUTHOR]

Changes in Vegetation Indices (VIs) over time can describe vegetation phenology; however, it is not known which phenological features contribute most to land cover classification. Feature selection could potentially solve this problem. In this study, phenological feature importance and selection were evaluated by using two-year Sentinel-2 (S-2) data and single-date PlanetScope (PS) to classify a 50 km2 native podocarp forest in New Zealand. The study area was classified into nine classes. Single-date PS and S-2 data were fused to a base image with the same spatial resolution as PS and 8 bands containing spectral data from S-2; this image was used to produce 30 Vegetation Indices (VIs). Phenological features – amplitude (AMP) and phase (PH) were extracted from these VIs using time-series S-2 only, and harmonic analysis in Google Earth Engine. For accurately classifying forests and identifying the most important features, three classification scenarios (fused bands & VIs, fused bands & phenological features, fused bands & VIs & phenological features) were developed using a Neural Network. Variable Selection Using Random Forest (VSURF) was applied on these scenarios to evaluate the impact of feature selection. Results indicate that VSURF could reduce the time needed for the classification while maintaining a comparable level of accuracy. Phenological features improved accuracy from 90% to 94%, driven mostly by Red-Edge Triangulated Vegetation Index-AMP&PH, Normalised Near-Infrared-PH, Greenness Index-PH, Water Body Index-PH, Normalised Difference Vegetation Index-PH, Normalised Green-PH, Red-Edge Normalised Difference Vegetation Index-PH, Leaf Chlorophyll Content-AMP, and Simple Near-Infrared and Blue Ratio-PH. These features reflect changes in the structure, biochemical, and physiological characteristics of the canopy. A lack of ground-based measurements precluded an evaluation of the accuracy of these phenological aspects and an explanation of their distinctive contribution to the model. Overall, the findings show that specific phenological features can improve the classification of New Zealand's indigenous podocarp forests. [ABSTRACT FROM AUTHOR]

Vegetation mapping from remote sensing data has proven useful for monitoring ecosystems at local, regional and global scales. Generally based on supervised classification methods, ecosystem mapping needs representative and consistent labelling. Such completeness is often difficult to achieve and requires the exclusion of minority species poorly represented in the studied scene in the training base. This exclusion leads to wrong predictions in the resulting map. In this study, the use of a posteriori classification rejection methods to limit the errors associated with minority species was evaluated in three different mapping scenarios: classification according to vegetation layers, prediction of genera from various vegetation types from low vegetation to trees and mapping of habitat (assemblages of species). For this purpose, several supervised classification methods based on Support Vector Machines (SVM), Random Forests (RF) and Regularized Logistic Regression (RLR) algorithms were first applied to hyperspectral images covering the reflective domain. On these classifications, the usual evaluation methods (confusion matrix and its derivatives calculated on an independent test set composed of the majority species) showed performances similar to those of the state-of-the-art. However, the introduction of a new score taking into account minority species demonstrated the need to include them in the evaluation process to provide robust performance quantification representing map effectiveness. Three a posteriori rejection methods, based on simple thresholding, K-means and SVM algorithms, were well suited to monitor minority species. The performance gain depended on the mapping scenario, the machine learning model and the rejection method. An increase in performance with the inclusion of minority species of up to 12% could be observed through the new score. These methods also detected a similar proportion of prediction errors associated with predominant species [ABSTRACT FROM AUTHOR]

Agricultural drought is a complex phenomenon with numerous consequences and negative implications for agriculture and food systems. The Sahel is frequently affected by severe droughts, leading to significant losses in agricultural yields. Consequently, assessing vulnerability to agricultural drought is essential for strengthening early warning systems. The aim of this study is to develop a new multivariate agricultural drought vulnerability index (MADVI) that combines static and dynamic factors extracted from satellite data. First, pixel temporal regression from 1981 to 2021 was applied to climatic and biophysical covariates to determine the gradients of trend magnitudes. Second, principal component analysis was applied to groups of factors that indicate the same type of vulnerability to configure the basic equation of vulnerability to agricultural drought. Then, random forest (RF), K-nearest neighbours (KNN), support vector machine (SVM) and naïve Bayes (NB) were used to predict drought vulnerability classes using the 28 factors as inputs and 708 pts of randomly distributed class labels. The results showed statistical agreement between the predicted MADVI spatial variability and the reference model (R=0.86 for RF) and its statistical relationships with the vulnerability subcomponents, with an R=0.73 with exposure to climate risk, R=0.64 with the socioeconomic sensitivity index, R=0.6 with the biophysical sensitivity index and a relatively weak correlation (R=0.21) with the physiographic sensitivity index. The overall vulnerability situation in the watershed is 21.8% extreme, 10% very high, 16.8% high, 27.7% moderate, 22.2% low and 1.5% relatively low considering the cartographic results of the predicted vulnerability classes with SVM having the best performance (accuracy=0.96, Kappa=0.95). The study is the first approach that uses the gradients of magnitudes of satellite covariate anomaly trends in multivariate modelling of vulnerability to agricultural drought. It can be easily scaled up across the Sahel region to improve early warning measures related to the impacts of agricultural drought. [ABSTRACT FROM AUTHOR]

Land cover (LC) classification remains a challenging task due to the diversity of terrain and topography, limited prior knowledge, and complex data sets. These variations can lead to differences in illumination and shading, which can affect the appearance of objects in the image. The local binary pattern (LBP) model is an effective technique for capturing local texture information in an image, which can help to overcome the effects of topography diversity by analysing patterns in different image regions, even if the illumination or shading conditions are different. However, LBP alone is inadequate for characterizing high-resolution remotely sensed images with complex semantic content as it only utilizes sign information in the local region. In this paper, a new texture characterization descriptor, known as completed homogeneous LBP (CHLBP), is proposed as an improved version of homogeneous LBP (HLBP) for LC classification of remotely sensed images. The CHLBP method mainly involves the following steps: first, sign and magnitude information from the HLBP descriptor is extracted, providing an effective alternative to the HLBP complementary contrast measure. Second, for each sign and magnitude function, a new splitting factor δ is used to obtain a depth relationship between the centre and its neighbouring pixels and to enhance noise robustness. Finally, the centre pixels representing the image grey level are also considered to contain discriminative information. The performance of our descriptor is evaluated using four challenging texture databases: Outex (TC10, TC12), Geofen Image Dataset (GID), and large-scale aerial (AID). Extensive experiments were performed using four classifiers: Support Vector Machine (SVM), K-Nearest Neighbours (KNN), Random Forest (RF), and Multi-Layer Perceptron (MLP), demonstrating the effectiveness and robustness of our descriptor against noise and free noise conditions in public remote sensing datasets. [ABSTRACT FROM AUTHOR]

Mapping and monitoring disturbances in vegetation over large areas demand reliable approaches and accurate end-user maps. Methods and algorithms have been developed to meet satisfactory disturbance map accuracies, and the combination of multiple approaches has shown promise as a reliable alternative to any single method. However, extracting meaningful disturbance information from these combined methods is still challenging. Data variance from environmental conditions and disturbance drivers leads to spatial-temporal heterogeneity in land surfaces over large areas, which results in mapping errors. We evaluated the effectiveness of ensemble classification and data-driven regionalization for mapping vegetation disturbances at a broad scale. Using Google's Earth Engine cloud computing platform, our ensemble approach combines multispectral LandTrendr outputs reflecting preliminary disturbance information in a Random Forest model to map disturbances in Minas Gerais, Brazil. We then applied an unsupervised clustering technique to perform data-driven regionalization of our study area using several sources of environmental and anthropogenic information and analysed gains and losses in map accuracies. Our results indicated gains in accuracy by the ensemble method compared to non-ensemble methods of disturbance mapping, which ranged from 7.3 to 29.9% in overall accuracy at the 5% significance level. Data-driven regionalization addressed complexities arising from variability in vegetation types, local climate, and topography across our study area, identifying climate and seasonal metrics as important variables for reducing uncertainties in vegetation disturbance maps. The integration of these techniques has revealed significant potential for increasing map accuracy and has provided important insights into the development of disturbance mapping methods in heterogeneous environments. [ABSTRACT FROM AUTHOR]

The quantification of grassland above-ground biomass (AGB) and grazing intensity (GI) and their distribution in space is of great significance to grassland management and eco-conservation. Remote-sensing technology is widely applied, but it is difficult to measure accurately when monitoring GI. In this study, the neural network, random forest and statistical function models of the relationship between Landsat NDVI and AGB were constructed by field survey and literature data collection in Inner Mongolia grassland, China. By comparing the accuracy among the three models, we constructed a remote-sensing retrieving model of grass AGB. We also estimated the grassland AGB during the peak growing season (August) for Inner Mongolia. Frequency histograms were then made to identify AGB thresholds under four GI levels (light or ungrazed, moderate grazing, overgrazing and extreme grazing) for each of three grassland types (meadow steppe, typical steppe and desert steppe). This study shows that the random forest model simulates grass AGB more accurately than other models. The spatial distribution of AGB in Inner Mongolia grasslands showed a tendency of decreasing from southeast to northwest, with an increasing trend in the last 10 years. The four GI levels in 2021 accounted for 18%, 25%, 36% and 21% of the grasslands in Inner Mongolia, respectively, and over the last 10 years the GI first improved and then deteriorated. This study provides a guideline to remote monitoring for grassland AGB and GI, and supplies scientific support for sustainable management and grassland restoration of large-scale grasslands. [ABSTRACT FROM AUTHOR]

Estimation of biophysical variables such as leaf area index (LAI) and canopy chlorophyll content (CCC) at high spatiotemporal resolution is important for managing natural and heterogeneous environments. However, accurate estimation of biophysical variables particularly over heterogeneous environments remains a challenge. The objective of the study was to develop locally parameterized grass LAI and CCC empirical models using the Sentinel-2 variables combined with the Stepwise multiple linear regression (SMLR) and Random forest (RF) at the Golden Gate Highlands National Park (GGHNP) and Marakele National Park (MNP) in South Africa. Results showed that in MNP, SMLR yielded better LAI estimation with root mean squared error (RMSE) of 0.67 m2.m−2 and mean adjusted error (MAE) of 0.54, explaining 48% of LAI variability, when bands and indices are combined. In contrast, RF gave better CCC estimation i.e. RMSE and MAE of 17.08 µg.cm−2 and 13.18 respectively, explaining about 40% of CCC variability with Sentinel-2 bands only. In GGHNP, the RF models provided the best estimates of both LAI and CCC compared to SMLR models. Furthermore, the CCC and LAI estimation models of GGHNP showed improved model accuracies when 50% and 75% of the MNP field samples were transferred to the GGHNP models. In contrast, the CCC and LAI estimation models of MNP showed a decline in model performance across all scenarios where the GGHNP field samples were transferred to the MNP models. These findings have significant implications for the development of locally parameterized types of models that can provide improved and consistent site-specific accurate estimates of grass biophysical parameters over heterogeneous environments. [ABSTRACT FROM AUTHOR]

The rapid expansion of exotic saltcedar along riparian corridors has dramatically altered the landscape structure and ecological function of riparian habitats in the western United States. The development of accurate and reproducible mapping methods with remote sensing plays an indispensable role in the timely monitoring of saltcedar, re-evaluating its ecological functions, and establishing effective control measures. The utmost challenge for achieving this goal is manifested as the lack of time series of remote sensing images to capture the saltcedar phenology adequately. To this end, the newly available Landsat 9 images, combined with its counterpart of Landsat 8, offer a precious opportunity to compensate for the temporal image shortage. To understand Landsat 9 in the saltcedar classification and to discover helpful information for its application, this study presents the first attempt to classify saltcedar using intra-annual Landsat 8 and Landsat 9 images. We adopted two machine learning algorithms, support vector machine (SVM) and random forest (RF), to compare the performance of Landsat 9 and Landsat 8 for intra-annual saltcedar classification. In addition, we investigated the respective contribution of each spectral band to the overall performance and identified the optimal time window for saltcedar classification. The results indicated that the difference in classification performance between Landsat 9 and Landsat 8 was insignificant. The shortwave infrared bands associated with both Landsat 8 & 9 have contributed most to the process of saltcedar identification. Image acquired in July, November, and December yielded better results than other months for saltcedar classification. It is concluded that Landsat 8 & 9 constellation has the potential to refine saltcedar classification accuracy on larger spatial and temporal scales. [ABSTRACT FROM AUTHOR]

Drought disasters significantly threaten the stability of terrestrial ecosystems in arid and semi-arid regions. Although the impact of drought on vegetation has been extensively researched, there is still a lack of understanding regarding the ecological impacts of drought on the heterogeneous vegetation in northern China's arid and semi-arid regions. To investigate the spatial patterns of vegetation and differences in response to drought climate, the Normalized Difference Vegetation Index (NDVI) and Standardized Precipitation Evapotranspiration Index (SPEI) were used to determine vegetation characteristics and the seasonal sensitivity of vegetation types to drought response, as well as to identify the factors driving the sensitivity differences. In the study area, 83.1% of the vegetation exhibited a significant positive correlation with the drought index. The maximum correlation coefficient between vegetation and drought was 0.62, demonstrating that the majority of vegetation was influenced by drought. It was observed that different vegetation types responded differently to drought at varying spatial and temporal scales. The main drivers of vegetation response to drought in arid and semi-arid zones were quantified by using the random forest algorithm. The results indicated that climate and soil factors are the primary limiting factors. These findings enhance our comprehension of arid and semi-arid areas, the effect of drought on different plant species, and the factors that affect vegetation's response to drought. [ABSTRACT FROM AUTHOR]

In existing rooftop extraction methods, either too few or too many features in high spatial resolution remote-sensing image (HSRRSI) are used, reducing the rooftop extraction accuracy. Accordingly, a rooftop extraction method for HSRRSI based on sparse representation (SR) is proposed in this work. The optimal segmentation parameters are first determined by the ratio of mean difference to neighbours to standard deviation index method and maximum area method. Thereafter, the optimal feature subset of HSRRSI is constructed on the basis of the L1 regularization SR model to remove redundant features. Finally, a random forest classifier is used to extract rooftops based on the optimal feature subset. Results show that the overall accuracy of the two study areas in Zhanggong District are 0.91776 and 0.88313, respectively. This study can help in effectively extracting rooftops from HSRRSI, which is of great significance in urban planning, population statistics and economic forecasting. [ABSTRACT FROM AUTHOR]

Accurate and reliable parcel-level apple orchard mapping is required for many precise agriculture application models, including planting suitability evaluation, standardized production, and personal agricultural operation loan approval. However, in hilly areas where smallholder management predominates, the highly fragmented and heterogeneous agricultural landscape means that fine parcel-level apple orchard mapping remains challenging. This paper proposes a parcel-level apple orchard mapping method based on feature-level spatiotemporal data fusion, which is suitable for hilly areas where smallholder management predominates. First, a hierarchical strategy that simulates human image cognition processing was used to extract redundant candidate parcels from a very high spatial resolution (VHR) image (Google Earth image with a spatial resolution of 0.6 m). Second, deep learning models, including a Depth-wise Asymmetric Bottleneck Network (DABNet) and long short-term memory (LSTM), were used to extract implicit spatial and time series features of the parcels. Third, the implicit features extracted by the deep learning models were formatted into meta-features, which then formed the feature space together with the morphological and geographical features of the parcel. Fourth, based on the constructed parcel feature space, a random forests (RF) model was used to classify candidate parcels. The experiment was carried out in the town of Guanli, southwest of Qixia city, Shandong Province, China: 21,123 apple orchard parcels were extracted from 31,235 candidate parcels. The overall accuracy (OA) of the parcel-level mapping result was 0.919. The parcel features were combined according to their types, and the performance of different feature combinations for parcel classification was further compared, demonstrating that the proposed meta-features had a stronger spatial information description capability than traditional features. Moreover, the mean decrease in the accuracy (MDA) index was used to evaluate the importance of each feature. And spatial-information-related meta-features were revealed to play the most important role in parcel classification. This method provides methodological references for parcel-level orchard mapping in hilly areas where smallholder management predominates and can be applied to improve the monitoring of orchards in such areas. [ABSTRACT FROM AUTHOR]

Maize is an important cereal crop and plays an important role in agriculture. The leaf area index (LAI) and the fraction of absorbed photosynthetically active radiation (FPAR) are important vegetation canopy structure parameters that are closely related to crop photosynthesis, transpiration and respiration and are of great significance for crop yield estimation. Hyperspectral data can obtain the spectral characteristic parameters of crops, and LiDAR data can obtain the vertical structure index. The former focuses on horizontal characteristic information, and the latter focuses on vertical characteristic information. In this paper, random forest method is used to carry out the inversion modelling of maize LAI and FPAR, and the effects of hyperspectral data resolution, LiDAR data density and scanning angle on the inversion accuracy are tested and analysed. The results show that there is good accuracy and stability of the random forest inversion model (LAI: 0.53

Oil spill mapping and detection represent a relevant issue from an environmental point of view, given the effects on marine ecosystems. This study presents a new feature space assessment protocol for oil spill mapping using the Google Earth Engine (GEE). First, we selected five free Sentinel-1A sensor images from the GEE catalogue. Next, we processed the features evaluated from Gray Level Co-occurrence Matrix (GLCM) spectral and texture data. A recursive protocol that comprises a sequential classification of the evaluated image was also applied, wherein each iteration, the feature with less importance, was removed based on the Gini index. We used the Random Forest algorithm for image classification. Each image was trained on 10,000 points and evaluated for accuracy, with an equal number of points collected independently. Our results showed that the Sum Average (Savg), Convolution Smooth (Smooth), Cluster Shade Shade, and Gray level Correlation (Corr) features were essential to identify oil spills and increase the accuracy values. The best classification results based on the features removal experiment and global accuracy were Angola (0.9960), Trinidad and Tobago (0.9829), Italy (0.9506), Kuwait (0.9547), and Dubai (0.9344). Furthermore, it revealed that the protocol created was essential for better understanding the parameter space to detect oil spills with SAR images. [ABSTRACT FROM AUTHOR]

The tree skeleton is a powerful tool for reasoning about the structure of forest, because of its ability to describe the topology and geometry of a tree in a simple and compact form. Terrestrial Laser Scanning (TLS) is one of the modern precision mapping techniques, it can provide high-precision point cloud data for digitization studies of forests. Accurate extraction of the stem and branch skeleton from TLS data is of great importance for studying the morphological structure and 3D visualization of trees. Although good progress has been made in the study of extracting the skeleton of 3D trees, there are still some problems, such as difficult to implement calculations on a large number of point clouds. Considering these problems, in this paper, cherry and begonia trees in a plot at Baima campus of Nanjing Forestry University are selected as the research object, and the space colonization algorithm is directly applied to the TLS point cloud data, but we find that the space colonization method would lead to over-extension of the skeletal trunk of the trees. Therefore, improvements are made on the basis of the original algorithm. Pre-processing the point cloud data with noise filtering and identifying the number of bifurcations using the DBSCAN clustering in advance makes the experimental results clearer visually. We select trees of different shapes within the plots for skeleton extraction and parameter validation was performed to illustrate the effectiveness of the improved method. The proposed method extracts the skeleton in the point cloud of 100 reference trees. The comparative measurement experiments with real values were performed, the root mean square error (RMSE) for the branch angle, trunk length and branch length were 3.309 degree, 0.069 m, 0.051 m respectively. These results demonstrate that the proposed method can accurately extract the skeleton points of a single tree. [ABSTRACT FROM AUTHOR]

Soil organic carbon (SOC) is one of the key soil components for cultivated soils. SOC is regularly monitored and mapped to improve the quality, health, and productivity of the soil. However, traditional SOC-level monitoring is expensive for land managers and farmers. Estimating SOC using satellite imagery provides an easy, efficient, and cost-effective way to monitor surface SOC levels. The objective of this study was to estimate the surface SOC distribution in selected soils of Major Land Resource Areas (MLRA), 102A (Rolling Till Plain, Brookings County, SD), and 103 (Central Iowa and Minnesota Till Prairies, Lac qui Parle County, MN), using satellite imagery with different resolutions (Landsat 8 and PlanetScope). The dominant soils in the study area are Haplustolls, Calciustolls, and Endoaquolls, which are formed in silty sediments, local silty alluvium, and till. Landsat 8 and PlanetScope spectral bands were used to develop SOC prediction models. Parametric and data-driven methods were employed to predict the SOC. Multiple linear regression and Linear Spatial Mixed Model (LSMM) were used on the Landsat 8 and PlanetScope data. In addition to the parametric models, Regression Trees and Random Forest were also employed on both satellite data. The results showed that reduced LSMM provided the lowest RMSE, which are 0.401 and 0.367 for Landsat 8 and PlanetScope, respectively. Furthermore, the random forest has the highest RPD and RPIQ for Landsat 8 (RPD of 2.67 and RPIQ of 2.49) and PlanetScope (RPD of 2.85 and RPIQ of 3.7). In all cases, models obtained from PlanetScope are better than those obtained from Landsat 8. [ABSTRACT FROM AUTHOR]

One of the main applications of machine learning (ML) in remote sensing (RS) is the pixel-level classification of satellite images into land cover types. Although classes with different spectral signatures can be easily separated, e.g. aquatic and terrestrial land cover types, others have similar spectral signatures and are hard to separate using only the information within a single pixel. This work focused on the separation of two cover types with similar spectral signatures, cocoa agroforest and forest, over an area in Pará, Brazil. For this, we study the training and application of several ML algorithms on datasets obtained from a single composite image, a time-series (TS) composite obtained from the same location and by preprocessing the TS composite using simple TS preprocessing techniques. As expected, when ML algorithms are applied to a dataset obtained from a composite image, the median producer's accuracy (PA) and user's accuracy (UA) in those two classes are significantly lower than the median overall accuracy (OA) for all classes. The second dataset allows the ML models to learn the evolution of the spectral signatures over 5 months. Compared to the first dataset, the results indicate that ML models generalize better using TS data, even if the series are short and without any preprocessing. This generalization is further improved in the last dataset. The ML models are subsequently applied to an area with different geographical bounds. These last results indicate that, out of seven classifiers, the popular random forest (RF) algorithm ranked fourth, while XGBoost (×GB) obtained the best results. The best OA, as well as the best PA/UA balance, were obtained by performing feature construction using the M3GP algorithm and then applying XGB to the new extended dataset. [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]

Medium Resolution Spectral Imager-II (MERSI-II) on board FY-3D satellite is a visible and infrared spectral imaging instrument featured with six 250-m spatial resolution channels (blue, green, red, NIR, 10.8 µm, 12.0 µm). Such unique spectral and spatial attributes provide significant opportunities and challenges for land cover (LC) mapping research. This paper reports the production of a 250-m spatial resolution global LC classification (under IGBP scheme) from MERSI-II NDVI operational products from May 2019 to April 2020. To extract robust features for discriminating LC types, we adopted three classic compositing criteria (MaxNDVI, Mean8NDVI, and Mean4BT) and two new compositing criteria (3minRedBlue and 3minRedNIR) to obtain high-quality annual composites. Random Forest (RF) algorithm was used for supervised learning, and 10-fold cross validation (CV) based on 'site-splits' was implemented to evaluate classification accuracy. The derived MERIS-II LC map was compared at pixel level with two widely used LC products (MODIS and VIIRS), and consistency among the three maps was investigated. Results showed that the higher resolution afforded by 250-m MERSI-II data in combination with a refined input features produced a global LC map with an overall accuracy of 79–86%. Most classes exhibited low commission and omission errors except for some classes of mixed nature (i.e. wetland, cropland/natural vegetation mosaic, savanna, woody savanna). The MERSI-II LC map generally agreed better with MODIS than with VIIRS, and the proportion of land pixels (excluding Antarctic) for which MERSI-II shows the same LC label with either MODIS or VIIRS is very large (0.70–0.82). Composite data contributed differently to total accuracy, with Mean8NDVI being the most spatial coherent and most contributive and MaxNDVI being the opposite. The other three composites (Mean4BT, 3minRedBlue, and 3minRedNIR) showed moderate spatial coherence and contributed to classification variedly by extracting information from different phenology phases. The unique 250-m blue and green bands of MERSI-II were found to be helpful to discriminate between wetland and needleleaf forest in high latitudes and between cities and grassland. This study confirms that with the RF algorithm (one of the state-of-art supervised machine learning algorithms) and image compositing techniques, accurate LC map can be derived from MERSI-II. Future work will attempt to include more bands and compositing criteria to further improve classification accuracy. [ABSTRACT FROM AUTHOR]

The retrieval of forest variables from optical remote sensing data using physically-based models is an ill-posed problem and does not make full use of the high spatial resolution imagery that is becoming available globally. A possible solution to this is to use prior information about the retrieved variables, which constrains the possible solutions and reduces uncertainty in forest variable estimation. Therefore, we tried to quantify physically-based parameters that could be retrieved using the second-order statistics of measured and simulated very-high-resolution (pixel size less than 1 m) images of Finnish boreal forests. These forests have a well-defined structure and are usually not closed, i.e. the reflected signal has a considerable contribution from a green forest floor. We retrieved the second-order statistics using variograms and Fourier amplitude spectra. We found, in line with previous studies, that the range of variograms correlates well ( r = 0.83) with the mean crown diameter for spatially homogeneous forest patches, and it can be used to estimate crown diameters with reasonable accuracy (RMSE = 0.42 m). We present a novel approach, which uses the Fourier amplitude spectrum to study the spatial structure of a forest. The approach provided encouraging results with the measured data: despite the lower accuracy (RMSE = 0.67 m) compared with variograms, we found that it could also be used to estimate mean crown diameters for heterogeneous forest areas. The Fourier amplitude spectrum approach did not work with the simulated images. Our results highlight the possibility to obtain further information from very-high-resolution images of forests to solve the ill-posed problem of forest variable estimation from optical remote sensing data using physically-based models. [ABSTRACT FROM AUTHOR]

The Andean páramo is a biodiverse and vulnerable tropical high-mountain region, whose spatio-ecological patterns remain understudied. The lack of general characterization of its overall extent, land-cover classes, and treeline spatial features hinders our capacity to understand its responses to human impacts and predict future land-system changes. To address this knowledge gap, we classified the land-cover of the páramo in the Northern Andes. Moreover, we estimated (1) the páramo's total extent and its distribution among the Andean countries, (2) the relative extent of 12 of its main land-cover classes, categorized into natural vegetation, natural abiotic and anthropogenic groups, and (3) the preliminary position and anthropogenic influence of its bordering treeline. Relying on Landsat 8 imagery, we performed hybrid manual-automated classifications using the Maximum Likelihood and Random Forest algorithms. The two resulting final classifications were manually checked for errors compared to Google Earth and VegPáramo data, and used to produce the expert classification. Finally, we delimited the treeline based on regional forest connectivity, and applied it to the expert classification to evaluate páramo elevations, surface areas and land-cover classes above the Andean treeline. The páramo extent was estimated at 24,301 km2, distributed between Ecuador (47%), Colombia (43%), Venezuela (8%) and Peru (2%). Natural vegetation, especially shrublands, rosette plant communities and grasslands were dominant (altogether, 65%), whereas classes reflecting intense land-use covered 12% overall. The average treeline reached 3546 m and was bordered uphill at 16% with anthropogenic land-cover classes. The páramo's extent is smaller than previously suggested. It remains a (semi-) natural region, yet the expansion of crops and pastures towards high elevations is a critical concern for long-term sustainability. Future research can build on our findings to predict land-system changes and assess priority areas for conservation. We recommend for future research to focus on remnant forest patches and treeline connectivity in priority. [ABSTRACT FROM AUTHOR]

Spartina alterniflora has become one of the top exotic invasive plants in coastal China. Accurate and timely mapping of coastal wetlands is critical to understand the spatiotemporal dynamics of S. alterniflora invasion and native species degradation. Due to the humid weather and frequent cloud cover in coastal wetlands, pixel-based classifications based on time-series satellite observations have been increasingly utilized to mitigate this problem. However, the spatial and temporal variations in valid observations might affect pixel-based classification accuracies, while this issue has been paid little attention in previous research. In this paper, we conducted annual mapping of the Yellow River Delta (YRD) wetland from 2008 to 2019 based on time-series Landsat 5/7/8 images using harmonic regression analysis and random forest classification on Google Earth Engine. We further analysed how valid observations and temporal distribution of observations affected the classification accuracy. Our results showed that the overall classification accuracies ranged from 87.25% in 2012 to 92.36% in 2018. The mapping results showed that S. alterniflora expanded from 39.91 ha to 4672.38 ha from 2008 to 2019, encroaching seagrass beds with 902.32 ha and encroaching S.salsa and bare flat with 3730.15 ha. This is the first time that maps of coastal wetlands of YRD are produced at an annual time step. Our analysis showed that the number of valid observations affected classification accuracies both temporally and spatially. Years with higher density of observations witnessed higher overall accuracies compared to the years with fewer observations. Pixels with more observations had a higher chance to be correctly classified compared to the pixels with fewer observations. Harmonic regression features helped improve classification accuracies (increase in overall accuracies from 0.15% to 3.32%), especially for vegetation types (increase in F-score from 0.19% to 3.86%). Greater number of valid observations enhances the importance of harmonic regression features in classification. The critical months identified were March, July, and October, and the combinations of these months achieved better accuracies (91.24%) than using all other observations (89.68%) in 2018, suggesting that temporal distributions of observations could be more important than the number of observations for harmonic-based classification of coastal wetlands such as YRD. [ABSTRACT FROM AUTHOR]

Soybean is one of the most important crops in the world. There were dramatic changes in soybean plantation in China in recent decades. Heilongjiang Province has the greatest acreage of soybean plantation in China. Therefore, it is taken as a representative soybean region in this study for monitoring purposes. A 30-m dataset that mapped the areas of soybean planting in Heilongjiang Province was produced, for the period of 1984 ~ 2020, based on Landsat-5/7/8 images and the Google Earth Engine (GEE) platform. Maps were made with a random forest classifier and had overall accuracies better than 84%. The areas planted with soybean were calculated at both the provincial and city level and these were compared with the areas obtained from official statistics. At the provincial level, for more than 85% of the time periods, the areas given in the official statistics fell within the confidence intervals of the estimates. At the city level, soybean areas were also in good agreement with the statistical data in most years giving average R2 values as 0.75. These results demonstrate the effectiveness of using phenological features extracted by a double-logistic model and a linear harmonic model with the RF classifier to obtain a long time-series of soybean maps without using samples obtained by field measurements every year. [ABSTRACT FROM AUTHOR]

Canopy cover is an important parameter in forest ecosystems and has diverse applications in a wide variety of fields. However, estimation of forest canopy cover from high spatial resolution images is an arduous task. To evaluate the utility of high spatial resolution images for estimation of canopy cover of artificial forests, the performance of canopy cover estimation models across sensors and across sites was evaluated. Choosing Wangyedian Forest Farm and Gaofeng Forest Farm as experimental areas, based on Chinese high spatial resolution Gaofen-1 (GF-1) and Ziyuan-3 (ZY-3) satellite images, three models, namely multiple linear regression (MLR), generalized additive model (GAM), and Random Forest (RF) were established. The capabilities of the two data sources and three model types to estimate canopy cover of the forests were compared and the portability of the three models across sensors and sites was analysed. The ranges of the root mean square error (RMSE) and relative root mean square error (rRMSE) of the MLR, GAM, and RF models for each study area established from GF-1 and ZY-3 were 0.0632–0.1205 and 9.98–19.93%, respectively. GF-1 showed better performance than ZY-3 when using the same model type. Whether the objective was across sensors or across sites transplantation, the accuracy of the model decreased with transplantation of the model. The MLR model was non-significant when transplanted across sites and model migration failed, whereas across sensors and across sites transplantation of RF and GAM models was successful. The three evaluation indices of transplanted GAM models were significantly superior to those of transplanted RF and MLR models. Thus, the GAM model showed strong portability and the highest model stability. [ABSTRACT FROM AUTHOR]

This paper aims to provide a novel approach for improving urban land cover classification accuracy, which combines the Elephant Herding Optimization (EHO) algorithm as a meta-heuristic optimization method with the Random Forest (RF) classifier. The proposed approach involves both classifier hyperparameter tuning and feature selection for a data set of a selected urban area. The EHO and the RF algorithms were both utilized in a hybrid system (EHO-RF) to find an optimal classification model with a tuned set of hyperparameters and features (predictor variables). EHO-RF model tuning was followed by backward feature elimination using RF, which further reduced data dimensionality. This work also combines two well-known optimization and feature selection algorithms to enhance accuracy assessment of the proposed approach. Grid search optimization was combined with Variable Selection Using Random Forests (VSURF) algorithm to build an optimization system (Grid-VSURF) that is similar to the workflow of EHO-RF. Objective functions for both Grid-VSURF and EHO-RF were designed to perform 10-fold cross-validation to reduce over-fitting. An area in Deerfield Beach city, Florida, USA, was selected as the study area representing an urban land cover. The testing dataset used in this study represents a performance benchmarking dataset, which has been utilized in many studies. EHO-RF results showed a significant improvement with an overall map accuracy of 83.83%, compared with Grid-VSURF results with an overall map accuracy of 74.75%. The number of input features was significantly reduced by both approaches with 82.30% reduction using EHO-RF with backward elimination and 89.70% reduction using Grid-VSURF. [ABSTRACT FROM AUTHOR]

Historical CORONA imagery, a very high-resolution panchromatic imagery captured during 1960-72, has a lot of unexplored potentials to represent the land cover of the pre-Landsat era. This pre-Landsat land cover can be a valuable input to many long-term environmental modelling problems. While automated binary change detection and multi-class classification by manual photo-interpretation have been the two most popular applications of CORONA imagery, automated multi-class classification of CORONA imagery has not been explored. In this research, we propose two supervised classification methods capable of extracting spatial features for the classification of CORONA imagery. The first method utilizes two-dimensional convolutional neural networks (2D-CNN) for classification, whereas the second method involves stacking the CORONA imagery with its texture features obtained through geometric moments (GM) and performing pixel-wise classification using the random forest (RF) classifier and is termed as GM-RF. The proposed methods are tested on two study sites – the critical zone observatory (CZO) managed by the Indian Institute of Technology Kanpur in Uttar Pradesh (India) and a site near Chhapra in Bihar (India). For estimating the capabilities of the proposed methods and to compare the proposed methods with other methods, overall accuracy and Cohen's kappa coefficient (k) along with z-scores statistics are used. The effects of various hyperparameters (e.g. percentage of training data, patch size, order of geometric moments) on accuracy are also evaluated. The classification accuracies achieved through the 2D-CNN and GM-RF method in terms of ĸ are higher than 90% for both study sites indicating that the proposed methods can classify CORONA imagery with little domain knowledge about the study areas, and hence can be used in a variety of applications. [ABSTRACT FROM AUTHOR]