Multi-temporal Sentinel-2 Images Research Articles

Comparison of Agricultural Crop Type Classifications with Different Machine Learning Algorithms (RF-SVM-ANN-XGBoost) by Generating Ground Truth Data from Farmer Declaration Parcels

In large-scale agricultural crop classification studies (Turkey, Adana, Çukurova Plain, 2500 km²), collecting sufficient and accurate ground truth data is costly, time-consuming, and unsustainable. This study utilized parcels registered in the Farmer Registration System (FRS) as ground truth data. By analyzing time series EVI curves, discrepancies were identified between declared and actual crops. Erroneous parcels were eliminated, and the corrected data were used in the classification process.Using multi-temporal Sentinel-2 images from 2021, this study compared the performance of Random Forests (RF), Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Extreme Gradient Boosting (XGBoost) algorithms for classifying crops like citrus, cotton, maize, peanut, sunflower, watermelon, wheat, and double-crop combinations (e.g., wheat-cotton, wheat-maize). The classification utilized 121 features (11 images × 10 Sentinel-2 bands + EVI). XGBoost achieved the highest overall accuracy (92.14%), followed by RF (89.15%), SVM (86.14%), and ANN (85.48%).The EVI index proved critical, particularly in separating spectral curves of double crops. While single crops like cotton, maize, and wheat yielded high classification accuracy, double crops with overlapping phenological stages had lower accuracy. The study highlighted that crops at distinct phenological stages performed well across algorithms, whereas crops with similar stages struggled to achieve high accuracy.This method of using corrected farmer-declared parcels (FDP) as ground truth data demonstrated high classification performance across all algorithms, proving its reliability. The findings emphasize that FDP can effectively replace traditional field data collection, reducing costs and improving efficiency. This classification approach supports agricultural production monitoring, yield estimation, water resource analysis, and sustainable policy-making, serving as a robust tool for agricultural evaluation

Winter Wheat Mapping in Shandong Province of China with Multi-Temporal Sentinel-2 Images

Wheat plays an important role in China’s and the world’s food supply, and it is closely related to economy, culture and life. The spatial distribution of wheat is of great significance to the rational planning of wheat cultivation areas and the improvement of wheat yield and quality. The current rapid development of remote sensing technology has greatly improved the efficiency of traditional agricultural surveys. The extraction of crop planting structure based on remote sensing images and technology is a popular topic in many researches. In response to the shortcomings of traditional methods, this research proposed a method based on the fusion of the pixel-based and object-oriented methods to map the spatial distribution of winter wheat. This method was experimented and achieved good results within Shandong Province. The resulting spatial distribution map of winter wheat has an overall accuracy of 92.2% with a kappa coefficient of 0.84. The comparison with the actual situation shows that the accuracy of the actual recognition of winter wheat is higher and better than the traditional pixel-based classification method. On this basis, the spatial pattern of winter wheat in Shandong was analyzed, and it was found that the topographic undulations had a great influence on the spatial distribution of wheat. This study vividly demonstrates the advantages and possibilities of combining pixel-based and object-oriented approaches through experiments, and also provides a reference for the next related research. Moreover, the winter wheat map of Shandong produced in this research is important for yield assessment, crop planting structure adjustment and the rational use of land resources.

JAXA’s new high-resolution land use land cover map for Vietnam using a time-feature convolutional neural network

Land use land cover (LULC) maps are crucial for various applications, such as disaster management, natural resource conservation, biodiversity evaluation, climate modeling, etc. The Japan Aerospace Exploration Agency (JAXA) has released several high-resolution LULC maps for national and regional scales. Vietnam, due to its rich biodiversity and cultural diversity, is a target country for the production of high-resolution LULC maps. This study introduces a high-resolution and high-accuracy LULC map for Vietnam, utilizing a CNN approach that performs convolution over a time-feature domain instead of the typical geospatial domain employed by conventional CNNs. By using multi-temporal data spanning 6 seasons, the produced LULC map achieved a high overall accuracy of 90.5% ± 1.2%, surpassing other 10-meter LULC maps for Vietnam in terms of accuracy and/or the ability to capture detailed features. In addition, a straightforward and practical approach was proposed for generating cloud-free multi-temporal Sentinel-2 images, particularly suitable for cloudy regions. This study marks the first implementation of the time-feature CNN approach for the creation of a high-accuracy LULC map in a tropical cloudy country.

Sentinel-2 versus PlanetScope Images for Goldenrod Invasive Plant Species Mapping

A proliferation of invasive species is displacing native species, occupying their habitats and degrading biodiversity. One of these is the invasive goldenrod (Solidago spp.), characterized by aggressive growth that results in habitat disruption as it outcompetes native plants. This invasiveness also leads to altered soil composition through the release of allelopathic chemicals, complicating control efforts and making it challenging to maintain ecological balance in affected areas. The research goal was to develop methods that allow the analysis of changes in heterogeneous habitats with high accuracy and repeatability. For this reason, we used open source classifiers Support Vector Machine (SVM), Random Forest (RF), and satellite images of Sentinel-2 (free) and PlanetScope (commercial) to assess their potential in goldenrod classification. Due to the fact that invasions begin with invasion footholds, created by small patches of invasive, autochthonous plants and different land cover patterns (asphalt, concrete, buildings) forming heterogeneous areas, we based our studies on field-verified polygons, which allowed the selection of randomized pixels for the training and validation of iterative classifications. The results confirmed that the optimal solution is the use of multitemporal Sentinel-2 images and the RF classifier, as this combination gave F1-score accuracy of 0.92–0.95 for polygons dominated by goldenrod and 0.85–0.89 for heterogeneous areas where goldenrod was in the minority (mix class; smaller share of goldenrod in canopy than autochthonous plants). The mean decrease in the accuracy analysis (MDA), indicating an informativeness of individual spectral bands, showed that Sentinel-2 bands coastal aerosol, NIR, green, SWIR, and red were comparably important, while in the case of PlanetScope data, the NIR and red were definitely the most important, and remaining bands were less informative, and yellow (B5) did not contribute significant information even during the flowering period, when the plant was covered with intensely yellow perianth, and red-edge, coastal aerosol, or green II were much more important. The maximum RF classification values of Sentinel-2 and PlanetScope images for goldenrod are similar (F1-score > 0.9), but the medians are lower for PlanetScope data, especially with the SVM algorithm.

Comparison between U-shaped structural deep learning models to detect landslide traces

Landslides endanger lives and public infrastructure in mountainous areas. Monitoring landslide traces in real-time is difficult for scientists, sometimes costly and risky because of the harsh terrain and instability. Nowadays, modern technology may be able to identify landslide-prone locations and inform locals for hours or days when the weather worsens. This study aims to propose indicators to detect landslide traces on the fields and remote sensing images; build deep learning (DL) models to identify landslides from Sentinel-2 images automatically; and apply DL-trained models to detect this natural hazard in some particular areas of Vietnam. Nine DL models were trained based on three U-shaped architectures, including U-Net, U2-Net, and U-Net3+, and three options of input sizes. The multi-temporal Sentinel-2 images were chosen as input data for training all models. As a result, the U-Net, using an input image size of 32 × 32 and a performance of 97 % with a loss function of 0.01, can detect typical landslide traces in Vietnam. Meanwhile, the U-Net (64 × 64) can detect more considerable landslide traces. Based on multi-temporal remote sensing data, a different case study in Vietnam was chosen to see landslide traces over time based on the trained U-Net (32 × 32) model. The trained model allows mountain managers to track landslide occurrences during wet seasons. Thus, landslide incidents distant from residential areas may be discovered early to warn of flash floods.

Contrastive-Learning-Based Time-Series Feature Representation for Parcel-Based Crop Mapping Using Incomplete Sentinel-2 Image Sequences

Parcel-based crop classification using multi-temporal satellite optical images plays a vital role in precision agriculture. However, optical image sequences may be incomplete due to the occlusion of clouds and shadows. Thus, exploring inherent time-series features to identify crop types from incomplete optical image sequences is a significant challenge. This study developed a contrastive-learning-based framework for time-series feature representation to improve crop classification using incomplete Sentinel-2 image sequences. Central to this method was the combined use of inherent time-series feature representation and machine-learning-based classifications. First, preprocessed multi-temporal Sentinel-2 satellite images were overlaid onto precise farmland parcel maps to generate raw time-series spectral features (with missing values) for each parcel. Second, an enhanced contrastive learning model was established to map the raw time-series spectral features to their inherent feature representation (without missing values). Thirdly, eXtreme Gradient-Boosting-based and Long Short-Term Memory-based classifiers were applied to feature representation to produce crop classification maps. The proposed method is further discussed and validated through parcel-based time-series crop classifications in two study areas (one in Dijon of France and the other in Zhaosu of China) with multi-temporal Sentinel-2 images in comparison to the existing methods. The classification results, demonstrating significant improvements greater than 3% in overall accuracy and 0.04 in F1 scores over comparison methods, indicate the effectiveness of the proposed contrastive-learning-based time-series feature representation for parcel-based crop classification utilizing incomplete Sentinel-2 image sequences.

Assessing spatial variations in soil organic carbon and C:N ratio in Northeast China's black soil region: Insights from Landsat-9 satellite and crop growth information

Monitoring soil properties, especially soil organic carbon (SOC) and Carbon-to-Nitrogen (C:N) ratios, is vital for understanding degradation and developing black soil conservation strategies. Most soil mapping research has leveraged satellite imagery from limited bare-soil periods. The predictive accuracy using time-series images capturing crop growth is underexplored. Moreover, while the new Landsat-9 satellite surpasses Landsat-8, its use in soil mapping is mostly uncharted. In this study, we compared the performance of single-date bare soil imagery (Landsat-8, Landsat-9, and Sentinel-2) and multi-temporal images (Landsat-8 and Landsat-9 combined, and Sentinel-2) using three machine learning techniques (Boosted Regression Tree, BRT; Random Forest, RF; Extreme Gradient Boosting, XGBoost) for mapping SOC and C:N ratio in a typical Northeast China black soil cropland region. The results revealed that single-date Landsat-9 exhibited great potential for soil mapping with the optimal XGBoost model, improving the prediction accuracy (in terms of R2) of SOC and C:N ratio by 15.01% and 30.07%, respectively, compared to Landsat-8, while also delivering performance slightly inferior to single-date Sentinel-2. Moreover, predictors derived from multi-temporal images significantly outperformed those derived from single-date images. The XGBoost models that utilized multi-temporal Sentinel-2 predictors achieved the highest prediction accuracy for both SOC (R2 = 0.676, RMSE = 1.928 g/kg, MAE = 1.580 g/kg, RPD = 1.535) and C:N ratio (R2 = 0.713, RMSE = 0.585, MAE = 0.484, RPD = 1.718). Interestingly, the combination of Landsat-8 and Landsat-9 demonstrated similar prediction accuracy but lower uncertainties in SOC and C:N ratio mapping compared to multi-temporal Sentinel-2 images. In addition, a comparison of the contemporary prediction soil maps with historical soil data revealed a continual decrease in SOC content and C:N ratio, suggesting a concerning trajectory that is detrimental to organic matter accumulation. Overall, this study highlights the efficacy of Landsat-9 and multi-temporal imagery incorporating crop growth information in accurately predicting soil properties and assessing their spatial variability.

The application of ResU-net and OBIA for landslide detection from multi-temporal Sentinel-2 images

ABSTRACT Landslide detection is a hot topic in the remote sensing community, particularly with the current rapid growth in volume (and variety) of Earth observation data and the substantial progress of computer vision. Deep learning algorithms, especially fully convolutional networks (FCNs), and variations like the ResU-Net have been used recently as rapid and automatic landslide detection approaches. Although FCNs have shown cutting-edge results in automatic landslide detection, accuracy can be improved by adding prior knowledge through possible frameworks. This study evaluates a rule-based object-based image analysis (OBIA) approach built on probabilities resulting from the ResU-Net model for landslide detection. We train the ResU-Net model using a landslide dataset comprising landslide inventories from various geographic regions, including our study area and test the testing area not used for training. In the OBIA stage, we first calculate land cover and image difference indices for pre-and post-landslide multi-temporal images. Next, we use the generated indices and the resulting ResU-Net probabilities for image segmentation; the extracted landslide object candidates are then optimized using rule-based classification. In the result validation section, the landslide detection of the proposed integration of the ResU-Net with a rule-based classification of OBIA is compared with that of the ResU-Net alone. Our proposed approach improves the mean intersection-over-union of the resulting map from the ResU-Net by more than 22%.

Deep Learning Application for Crop Classification via Multi-Temporal Remote Sensing Images

The combination of multi-temporal images and deep learning is an efficient way to obtain accurate crop distributions and so has drawn increasing attention. However, few studies have compared deep learning models with different architectures, so it remains unclear how a deep learning model should be selected for multi-temporal crop classification, and the best possible accuracy is. To address this issue, the present work compares and analyzes a crop classification application based on deep learning models and different time-series data to exploit the possibility of improving crop classification accuracy. Using Multi-temporal Sentinel-2 images as source data, time-series classification datasets are constructed based on vegetation indexes (VIs) and spectral stacking, respectively, following which we compare and evaluate the crop classification application based on time-series datasets and five deep learning architectures: (1) one-dimensional convolutional neural networks (1D-CNNs), (2) long short-term memory (LSTM), (3) two-dimensional-CNNs (2D-CNNs), (4) three-dimensional-CNNs (3D-CNNs), and (5) two-dimensional convolutional LSTM (ConvLSTM2D). The results show that the accuracy of both 1D-CNN (92.5%) and LSTM (93.25%) is higher than that of random forest (~ 91%) when using a single temporal feature as input. The 2D-CNN model integrates temporal and spatial information and is slightly more accurate (94.76%), but fails to fully utilize its multi-spectral features. The accuracy of 1D-CNN and LSTM models integrated with temporal and multi-spectral features is 96.94% and 96.84%, respectively. However, neither model can extract spatial information. The accuracy of 3D-CNN and ConvLSTM2D models is 97.43% and 97.25%, respectively. The experimental results show limited accuracy for crop classification based on single temporal features, whereas the combination of temporal features with multi-spectral or spatial information significantly improves classification accuracy. The 3D-CNN and ConvLSTM2D models are thus the best deep learning architectures for multi-temporal crop classification. However, the ConvLSTM architecture combining recurrent neural networks and CNNs should be further developed for multi-temporal image crop classification.

Optimal Time Phase Identification for Apple Orchard Land Recognition and Spatial Analysis Using Multitemporal Sentinel-2 Images and Random Forest Classification

The significance of identifying apple orchard land and monitoring its spatial distribution patterns is increasing for precise yield prediction and agricultural sustainable development. This study strived to identify the optimal time phase to efficiently extract apple orchard land and monitor its spatial characteristics based on the random forest (RF) classification method and multitemporal Sentinel-2 images. Firstly, the Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Ratio Vegetation Index (RVI), and Difference Vegetation Index (DVI) between apple orchard land and other green vegetation (other orchards, forest and grassland) during the growing stage were calculated and compared to identify the optimal time phase for apple orchard land extraction; the RF classifier was then constructed using multifeature variables on Google Earth Engine to efficiently identify apple orchard land, and the support vector machine (SVM) classification results were used as a comparison; GIS spatial analysis, a slope calculation model, and Moran’s I and Getis-Ord GI* analysis were employed to further analyze the spatial patterns of the apple orchard land. The results found the following: (1) April, May, and October were the optimal time phases for apple orchard identification. (2) The RF-based method combining coefficients of indexes, the grayscale co-occurrence matrix, and 70% of the ground reference data can precisely classify apple orchards with an overall accuracy of 90% and a Kappa coefficient of 0.88, increasing by 9.2% and 11.4% compared to those using the SVM. (3) The total area of apple orchard land in the study area was 485.8 km2, which is 0.6% less than the government’s statistical results. More than half (55.7%) of the apple orchard land was distributed on the gentle slope (Grade II, 6–15°) and the flat slope (Grade I, 0–5°); SiKou, Songshan, and Shewopo contained more than 50% of the total orchard land area. (4) The distribution of apple orchard land has a positive spatial autocorrelation (0.309, p = 0.000). High–High cluster types occurred mainly in Sikou (60%), High–Low clusters in Songshan (40%), Low–High clusters in Sikou (47.5%), and Low–Low clusters in Taocun and Tingkou (37.4%). The distribution patterns of cold and hot spots converged with those of the Local Moran Index computation results. The findings of this study can provide theoretical and methodological references for orchard land identification and spatial analysis.

Mapping abandoned cropland using Within-Year Sentinel-2 time series

Against the background of the COVID-19 pandemic and various armed conflicts, the world is experiencing an unprecedented food crisis. The reclamation of abandoned cropland with food production potential may increase the global food supply in a short period of time, ensuring food security. At present, the extraction of abandoned cropland is mainly based on low- and medium-resolution remote sensing image data, making it difficult to extract fragmented areas in mountainous regions and to distinguish between abandoned cropland and transitional classes (such as fallow cropland). We developed a change-detection method based on within-year Sentinel-2 time series to extract cropland abandoned from 2018 to 2021 and defined four types of croplands, namely spontaneously abandoned, induced abandoned, fallow, and lost cropland, using Linxia County in mountainous China as the study region. First, cropland objects were generated from multi-temporal Sentinel-2 images using the multi-resolution segmentation method, and the land use map of Linxia County from 2017 to 2021 was drawn using random forest classifier. Second, through defining and identifying different cropland types, the interannual dynamic changes in cropland from 2018 to 2021 were extracted by analyzing the annual land use change trajectory. Third, by analyzing the normalized difference vegetation index (NDVI) time series of cropland within-year, the active and cultivated cropland sites within-year were extracted by threshold segmentation. Finally, the changes in the four cropland types were extracted by intersecting the two result types. Our method captured the object level changes well (overall mapping accuracy = 93 ± 5 %), and the extraction accuracy of abandoned cropland reached 81 ± 2 %. Abandoned cropland was mostly located in areas of medium quality and with a moderate distance from rural settlements. Reclamation can potentially increase the grain production in Linxia County by at least 3.6 % and needs to be combined with the local natural geography and human activities. Our method is a robust method for extracting abandoned cropland and may be applied to other research related to land use change.

Quantifying Temperate Forest Diversity by Integrating GEDI LiDAR and Multi-Temporal Sentinel-2 Imagery

Remotely sensed estimates of forest diversity have become increasingly important in assessing anthropogenic and natural disturbances and their effects on biodiversity under limited resources. Whereas field inventories and optical images are generally used to estimate forest diversity, studies that combine vertical structure information and multi-temporal phenological characteristics to accurately quantify diversity in large, heterogeneous forest areas are still lacking. In this study, combined with regression models, three different diversity indices, namely Simpson (λ), Shannon (H′), and Pielou (J′), were applied to characterize forest tree species diversity by using GEDI LiDAR data and Sentinel-2 imagery in temperate natural forest, northeast China. We used Mean Decrease Gini (MDG) and Boosted Regression Tree (BRT) to assess the importance of certain variables including monthly spectral bands, vegetation indices, foliage height diversity (FHD), and plant area index (PAI) of growing season and non-growing seasons (68 variables in total). We produced 12 forest diversity maps on three different diversity indices using four regression algorithms: Support Vector Machines (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), and Lasso Regression (LR). Our study concluded that the most important variables are FHD, NDVI, NDWI, EVI, short-wave infrared (SWIR) and red-edge (RE) bands, especially in the growing season (May and June). In terms of algorithms, the estimation accuracies of the RF (averaged R2 = 0.79) and SVM (averaged R2 = 0.76) models outperformed the other models (R2 of KNN and LR are 0.68 and 0.57, respectively). The study demonstrates the accuracy of GEDI LiDAR data and multi-temporal Sentinel-2 images in estimating forest diversity over large areas, advancing the capacity to monitor and manage forest ecosystems.

Extracting Tea Plantations from Multitemporal Sentinel-2 Images Based on Deep Learning Networks

Tea is a special economic crop that is widely distributed in tropical and subtropical areas. Timely and accurate access to the distribution of tea plantation areas is crucial for effective tea plantation supervision and sustainable agricultural development. Traditional methods for tea plantation extraction are highly dependent on feature engineering, which requires expensive human and material resources, and it is sometimes even difficult to achieve the expected results in terms of accuracy and robustness. To alleviate such problems, we took Xinchang County as the study area and proposed a method to extract tea plantations based on deep learning networks. Convolutional neural network (CNN) and recurrent neural network (RNN) modules were combined to build an R-CNN model that can automatically obtain both spatial and temporal information from multitemporal Sentinel-2 remote sensing images of tea plantations, and then the spatial distribution of tea plantations was predicted. To confirm the effectiveness of our method, support vector machine (SVM), random forest (RF), CNN, and RNN methods were used for comparative experiments. The results show that the R-CNN method has great potential in the tea plantation extraction task, with an F1 score and IoU of 0.885 and 0.793 on the test dataset, respectively. The overall classification accuracy and kappa coefficient for the whole region are 0.953 and 0.904, respectively, indicating that this method possesses higher extraction accuracy than the other four methods. In addition, we found that the distribution index of tea plantations in mountainous areas with gentle slopes is the highest in Xinchang County. This study can provide a reference basis for the fine mapping of tea plantation distributions.

Estimating Suspended Sediment Fluxes from the Largest Glacial Lake in Svalbard to Fjord System Using Sentinel-2 Data: Trebrevatnet Case Study

Glacier-fed hydrological systems in high latitude regions experience high seasonal variation in meltwater runoff. The peak in runoff usually coincides with the highest air temperatures which drive meltwater production. This process is often accompanied by the release of sediments from within the glacier system that are transported and suspended in high concentrations as they reach the proglacial realm. Sediment-laden meltwater is later transported to the marine environment and is expressed on the surface of fjords and coastal waters as sediment plumes. Direct monitoring of these processes requires complex and time-intensive fieldwork, meaning studies of these processes are rare. This paper demonstrates the seasonal dynamics of the Trebrevatnet lake complex and evolution of suspended sediment in the lake and sediment plumes in the adjacent Ekmanfjorden. We use the Normalized Difference Suspended Sediment Index (NDSSI) derived from multi-temporal Sentinel-2 images for the period between 2016–2021. We propose a new SSL index combining the areal extent of the sediment plume with the NDSSI for quantification of the sediment influx to the marine environment. The largest observed sediment plume was recorded on 30 July 2018 and extended to more than 40 km2 and a SSL index of 10.4. We identified the greatest sediment concentrations in the lake in the beginning of August, whereas the highest activity of the sediment plumes is concentrated at the end of July. The temporal pattern of these processes stays relatively stable throughout all ablation seasons studied. Sediment plumes observed with the use of optical satellite remote sensing data may be used as a proxy for meltwater runoff from the glacier-fed Trebrevatnet system. We have shown that remote-sensing-derived suspended sediment indexes can (after proper in situ calibration) serve for large scale quantification of sediment flux to fjord and coastal environments.

Açık kaynak kodlu Eo-learn kütüphanesi ve çok zamanlı Sentinel-2 görüntüleri ile tarımsal ürün sınıflandırması

In this study, agricultural crop classification for 2020 was carried out in Çivril-Baklan Plain, which is located between the borders of Denizli Province, Baklan, Çal and Çivril districts. The open-source Eo-Learn library that uses machine learning and deep learning algorithms in remote sensing studies and multi-temporal Sentinel-2 images was utilized in the classification process. In this study, the parcels registered in the Farmer Registration System (FRS) were used as reference parcels and before using FRS data as ground truth data, pre-editing and rule-based deletion processes were performed. By using Light Gradient Boosting Machines (LightGBM) algorithm, agricultural product pattern classification was carried out including cereal, maize, sugar beet, sunflower, hash, vineyard, fruit tree and clover crops. Classification results were evaluated using k-fold cross-validation with an overall accuracy of %93.5. A second accuracy assessment was performed with Agricultural Insurance Parcels (TARSİM) that were not included in the classification process as training data, achieving an overall accuracy of %91.1 and Kappa coefficient of 0.89.

Spatiotemporal monitoring of low water turbidity in Moroccan coastal lagoon using Sentinel-2 data

Water turbidity is a key index used to evaluate coastal water quality conditions. To determine the turbidity in large areas, remote sensing technologies have the potential to provide a cost-effective alternative to experimental methods that require extensive time for field data collection and laboratory analysis. This research focuses on the development of an accurate model to estimate and monitor low water turbidity using Sentinel-2 images and field measurements in the Sidi Moussa lagoon of the Moroccan Atlantic coast. Also, the effect of various factors was mentioned and compared qualitatively to the water turbidity variations, including spatiotemporal variation, wind, depth, geomorphology, and vegetation. Among all the regressions analyzed, the Red band was the most successful remote sensing information for water turbidity with a good validation accuracy of R2 = 0.87. ANOVA analysis showed that measured turbidity values vary differently from upstream to downstream of the lagoon. The analysis of the seasonal effect on the water turbidity using one-year multitemporal Sentinel-2 images showed that in the dry season, the mean turbidity is lower than in the rainy season. The water turbidity map derived using the developed algorithm can be used to support wetland management, the conservation of natural resources, and the selection of aquaculture areas. These results highlight the potential contribution of Sentinel-2 to map and monitor the low water turbidity.

Mapping tree species in natural and planted forests using Sentinel-2 images

ABSTRACT Using remote-sensing technology to accurately map the composition and distribution of tree species is vital for sustainable forest resource management. Sentinel-2 data with the dense time-series observations enable to identify tree species. However, few studies clarify the differences in classification using Sentinel-2 images in natural forest and planted forest. Two study areas with different forest environments (planted forest and natural forest) were selected to evaluate the potential of Sentinel-2 imagery. Our results show that red-edge band, short-wave infrared (SWIR) band and vegetation indices (VIs), such as Red-Edge Normalized Difference Vegetation Index (ReNDVI), Soil-Adjusted Vegetation Index (SAVI), and Enhanced Vegetation Index (EVI), have important effects on tree species classification, especially in the growing season. By using two machine learning algorithms (support vector machine [SVM] and random forest [RF]), the results show that the classification accuracy for planted forests (91.27% in SVM and 88.35% in RF) significantly exceeds that of natural forests (84.34% in SVM and 81.03% in RF). This accuracy difference may be related to the spatial heterogeneity inside the forest and the surrounding environmental implications. Although the multi-temporal Sentinel-2 images produce satisfactory accuracy for classifying tree species, further research is needed to improve the classification accuracy.

Identification of Soil Texture Classes Under Vegetation Cover Based on Sentinel-2 Data With SVM and SHAP Techniques

Understanding the spatial variability of soil texture classes is essential for agricultural management and environment sustainability. Sentinel-2 data offer valuable vegetation information as proxies for soil properties inference. However, the applications of them in soil texture classification are still limited. This study investigated the usefulness of Sentinel-2 data for predicting soil texture class using an interpretable machine learning (ML) strategy. Specifically, multitemporal Sentinel-2 images were used to get exhaustive vegetation cover information. Basic digital elevation map (DEM) derivatives and stratum were extracted. Three support vector machines with different input parameters (purely DEM derivatives and stratum, purely Sentinel-2, and Sentinel-2 plus DEM derivatives and stratum) were developed. Moreover, in order to improve the transparency in black box ML models, the novel SHapley Additive exPlanations (SHAP) method was applied to interpret the outputs and analyze the importance of individual variables. Results showed that the model with all variables provided desirable performance with overall accuracy of 0.8435, F1-score of 0.835, kappa statistic of 0.7642, precision of 0.8388, recall of 0.8355, and area under the curve of 0.9451. The model with purely Sentinel-2 data performed much better than that with solely DEM derivatives and stratum. The contributions of Sentinel-2 data to explain soil texture class variability were about 17%, 41%, and 28% for sandy, loamy and clayey soils, respectively. The SHAP method visualized the decision process of ML and indicated that elevation, stratum, and red-edge factors were critical variables for predicting soil texture classes. This study offered much-needed insights into the applications of Sentinel-2 data in digital soil mapping and ML-assisted tasks.

Anomaly Detection and Classification in Multispectral Time Series Based on Hidden Markov Models

Monitoring agriculture from satellite remote sensing data, such as multispectral images, has become a powerful tool since it has demonstrated a great potential for providing timely and accurate knowledge of crops. Detecting anomalies in time series of multispectral remote sensing images for crop monitoring is generally performed using a large sample of historical data at a pixel level. Conversely, this article presents a framework for anomaly detection (AD), localization, and classification that exploits the temporal information contained in a given season at a parcel level to detect and localize outliers using hidden Markov models (HMMs). Specifically, the AD part is based on the learning of HMM parameters associated with unlabeled normal data that are used in a second step to detect abnormal crop parcels referred to as anomalies. The learned HMM can also be used in time segments to temporally localize the anomalies affecting the crop parcels. The detected and localized anomalies are finally classified using a supervised classifier, e.g., based on support vector machines. The proposed framework is applicable to images partially covered by clouds and can handle a set of crop parcels acquired in the same season bypassing problems due to crop rotations. Numerical experiments are conducted on synthetic and real data, where the real data correspond to vegetation indices extracted from several multitemporal Sentinel-2 images of rapeseed crops. The proposed approach is compared to standard AD methods yielding better detection rates with the advantage of allowing anomalies to be localized and characterized.

Tropical Forests Classification Based on Weighted Separation Index from Multi-Temporal Sentinel-2 Images in Hainan Island

Tropical forests play a vital role in biodiversity conservation and the maintenance of sustainability. Although different time-series spatial resolution satellite images have provided opportunities for tropical forests classification, the complexity and diversity of vegetation types still pose challenges, especially for distinguishing different vegetation types. In this paper, we proposed a Spectro-Temporal Feature Selection (STFS) method based on the Weighted Separation Index (WSI) using multi-temporal Sentinel-2 data for mapping tropical forests in Jianfengling area, Hainan Province. The results showed that the tropical forests were classified with an overall accuracy of 93% and an F1 measure of 0.92 with multi-temporal Sentinel-2 data. As our results also revealed, the WSI based STFS method could be efficient in tropical forests classification by using a fewer feature subset compared with Variable Selection Using Random Forest (14 features and all 40 features, respectively) to achieve the same accuracy. The analysis also showed it was not advisable to only pursue a higher WSI value while ignoring the heterogeneity and diversity of features. This study demonstrated that the WSI can provide a new feature selection method for multi-temporal remote sensing image classification.