Sentinel-2 Imagery Research Articles

Understanding the Temporal Dynamics of Coherence and Backscattering Using Sentinel-1 Imagery for Crop-Type Mapping

Understanding the Temporal Dynamics of Coherence and Backscattering Using Sentinel-1 Imagery for Crop-Type Mapping

Dual-polarimetric method to detect burned area after the 2022 Stromboli Island wildfire

In this study, dual-polarimetric (DP) feature derived from Sentinel-1 Synthetic Aperture Radar (SAR) data is used to detect burned area (BA) related to the wildfire that hit Stromboli Island on May 2022. First, a DP change detection feature based on the difference of covariance matrices obtained from images collected before and after is exploited to detect the BA. Finally, the BA is extracted using a clustering algorithm based on k-means. To assess the performance of the proposed approach, the DP detector is contrasted with single polarimetric ones, and optical imagery from Sentinel-2 used as reference data for evaluating the accuracy of the DP method. Experimental results, obtained by processing a pair of C-band Sentinel-1 imagery collected before and after the 2022 Stromboli Island, wildfire shows that the proposed feature is able to detect the BA from the surroundings. In addition, the DP feature outperforms the SP ones.

Satellite observations of storm erosion and recovery of the Ebro Delta coastline, NE Spain

Storms and extremely energetic events may significantly impact the form and structure of beaches, and so cause erosive processes and coastal damages. Efficient management actions require an up-to-date and accurate knowledge of beach morphological changes, with the shoreline position being a good indicator of such changes. This work proposes the use of the open-source Shoreline Analysis and Extraction Tool (SAET) software for the definition of satellite-derived shorelines (SDSs) from L8 and Sentinel-2 imagery to reveal the shoreline position changes at the beaches of the Ebro Delta, NE Spain. Spatial-temporal models (STMs) of shoreline changes enable a characterisation of how the beaches responded to the storms of 2020. In conjunction with wave data, STMs enable an analysis of the erosive response to storm events, as well as a monitoring of subsequent beach recovery in the short and medium term (<1 year). Results show how Storm Gloria (January 2020, Hs max = 7.62 m) acted as a disruptive event and shifting point in the shoreline trend. As a response to that storm, major erosive processes occurred along the delta that caused an average shoreline retreat of 47 m. A progressive recovery during the spring and summer was mainly associated with periods of low wave energy. Nevertheless, by the end of the year a complete recovery had been achieved for about half of the coast, while the other half showed an average erosion of more than 10 m when compared to the pre-storm situation. Both the erosive and the recovery processes took place unevenly on different sections of the coast, probably dependent on factors such as the orientation of the beach and the pattern of longitudinal sediment transport along the coast.

Temporal development of the displacement field of the Ponzano landslide in February 2017

The conducted research determined the temporal evolution of the displacement field for the Ponzano landslide case study. The offset-tracking method, so far used mainly for the relatively rapid but uniform displacement of glaciers, was tested for the 2017 study of the Ponzano landslide in Italy. The suitability of the method for high-resolution TerraSAR-X and medium-resolution Sentinel-1 imagery was investigated. The results proved the applicability of the OT method for studying processes with high and variable displacement dynamics. However, for such purposes, high-resolution radar data are crucial. With an uncertainty in the determination of residual displacements of about ±1 m, it was shown that the values of residual displacements occurring up to several days after the main phase of landslide movements are within the range of uncertainty but are determinable. The research conducted in the paper filled a gap in the analysis of the phenomenon just after the main movement phase. It allowed determination of the time and speed of extinction of landslide movements.

Rapid wildfire damage estimation using integrated object-based classification with auto-generated training samples from Sentinel-2 imagery on Google Earth Engine

Satellite data are essential during wildfires for understanding its adverse effects and improving the effectiveness of rapid disaster management. However, existing techniques used for damage assessments are inaccurate and lack automation. In this study, we propose an integrated machine learning approach with auto-generated training samples for a rapid wildfire disaster response framework using Sentinel-2 imagery at 10 m resolution from Google Earth Engine (GEE). First, training samples of burned areas were obtained by utilizing textural data based on features that had changed because of the wildfire, and samples of unburned areas were obtained using the normalized difference vegetation index (NDVI). The images were categorized as burned and unburned images using the object-based image analysis (OBIA) classification method. Finally, using the classified maps, burn severity maps and estimated pixel counts for each severity class were generated and compared. The proposed method was implemented to put out a wildfire that broke out in Uljin, Gyeongsangbuk-do, South Korea in March 2022 and the transferability of the model was evaluated in Gangneung, Gangwon-do, South Korea. The study findings indicate that the random forest (RF) classifier acquired the greatest overall accuracy (OA) of 97.6 % in Uljin; additionally, the model transferability performed well in Gangneung with an OA of 93.8 %. The RF also generated the fewest pixels of the unchanged class when the burn severity map was evaluated. Overall, our study proposes a quick and automated approach for estimating wildfire damage that could be used for immediate mitigation actions.

Large-Scale Cotton Classification under Insufficient Sample Conditions Using an Adaptive Feature Network and Sentinel-2 Imagery in Uzbekistan

Cotton (Gossypium hirsutum L.) is one of the main crops in Uzbekistan, which makes a major contribution to the country’s economy. The cotton industry has played a pivotal role in the economic landscape of Uzbekistan for decades, generating employment opportunities and supporting the livelihoods of countless individuals across the country. Therefore, having precise and up-to-date data on cotton cultivation areas is crucial for overseeing and effectively managing cotton fields. Nonetheless, there is currently no extensive, high-resolution approach that is appropriate for mapping cotton fields on a large scale, and it is necessary to address the issues related to the absence of ground-truth data, inadequate resolution, and timeliness. In this study, we introduced an effective approach for automatically mapping cotton fields on a large scale. A crop-type mapping method based on phenology was conducted to map cotton fields across the country. This research affirms the significance of phenological metrics in enhancing the mapping of cotton fields during the growing season in Uzbekistan. We used an adaptive feature-fusion network for crop classification using single-temporal Sentinel-2 images and automatically generated samples. The map achieved an overall accuracy (OA) of 0.947 and a kappa coefficient (KC) of 0.795. This model can be integrated with additional datasets to predict yield based on the identified crop type, thereby enhancing decision-making processes related to supply chain logistics and seasonal production forecasts. The early boll opening stage, occurring approximately a little more than a month before harvest, yielded the most precise identification of cotton fields.

Monitoring lianas from space: Using Sentinel-2 imagery to observe liana removal in logged tropical forests

Liana removal – the cutting of over-abundant woody climbing plants (lianas) – has the potential to substantially increase tree growth and biomass accumulation across millions of hectares of degraded tropical forest. Satellite imagery could provide data capable of observing the effect of liana removal on the forest canopy, enabling the large-scale monitoring and validation of liana removal, which remains a key hurdle to its widespread implementation. Using a 320-ha liana removal experiment in Sabah, Malaysian Borneo, we tested whether a time series of Sentinel-2 images could observe the canopy signature of liana removal. Calculating a range of metrics derived from the Normalized Burn Ratio – a vegetation index based on spectral reflectance that differentiates leaf from non-leaf – we quantified satellite-derived canopy disturbance and fragmentation across a range of liana removal intensities and examined how canopy disturbance changed in the 12-months following removal treatments. We find that liana removal significantly increases canopy disturbance and fragmentation metrics one month after removal, with partial removal having a smaller effect than complete removal. The impact of liana removal on the canopy metrics declined over time, with measures of canopy disturbance and fragmentation largely indistinguishable from control forest within 12-months of treatment. Our findings evidence that freely available satellite imagery can be used to efficiently monitor large-scale liana removal applied at a range of intensities and suggest that partial liana removal could significantly reduce canopy disturbance of this restoration method.

Modeling strategies and influencing factors in retrieving canopy equivalent water thickness of mangrove forest with Sentinel-2 image

Canopy equivalent water thickness (EWT) is an essential indicator of plant water status related to plant growth, temperature maintenance and transpiration. To our knowledge, this study was the first to map the mangrove canopy EWT using Sentinel-2 image at the reserve scale, aiming to explore three modeling strategies in retrieving mangrove canopy EWT, including machine learning models (Random Forest Regression (RFR) and Adaptive Boosting (Adaboost)), radiative transfer models (RTMs, PROSAIL-D and biophysical processor in Sentinel application platform) and hybrid models (PROSAIL-D + RFR and PROSAIL-D + Adaboost). We further investigated the impacts of four ecogeographical factors (species distribution, slope, elevation and distance to dam) on the spatial distribution of canopy EWT using Geodetector method. The results showed that Adaboost (R2 = 0.593, RMSE = 0.771 kg/m2 and RPIQ = 1.820), PROSAIL-D (R2 = 0.451, RMSE = 0.937 kg/m2 and RPIQ = 1.537) and PROSAIL-D + Adaboost (R2 = 0.501, RMSE = 0.890 kg/m2 and RPIQ = 0.887) was the optimal machine learning, RTM and hybrid model in retrieving mangrove canopy EWT, respectively. The red-edge3, NIR and SWIR bands of Sentinel-2 imagery were sensitive to canopy EWT. Moreover, Geodetector analysis demonstrated that species distribution had the most significant impact on canopy EWT, with the interaction between species and distance to dams contributing the most to the spatial difference of canopy EWT. In conclusion, this study suggests that Adaboost model and the hybrid method of PROSAIL-D and Adaboost were promising to map large-scale canopy water conditions in mangrove ecosystems using Sentinel-2 imagery.

Mapping the tidal marshes of coastal Virginia: a hierarchical transfer learning approach

ABSTRACT Coastal wetlands, especially tidal marshes, play a crucial role in supporting ecosystems and slowing shoreline erosion. Accurate and cost-effective identification and classification of various marsh types, such as high and low marshes, are important for effective coastal management and conservation endeavors. However, mapping tidal marshes is challenging due to heterogeneous coastal vegetation and dynamic tidal influences. In this study, we employ a deep learning segmentation model to automate the identification and classification of tidal marsh communities in coastal Virginia, USA, using seasonal, publicly available satellite and aerial images. This study leverages the combined capabilities of Sentinel-2 and National Agriculture Imagery Program (NAIP) imagery and a UNet architecture to accurately classify tidal marsh communities. We illustrate that by leveraging features learned from data abundant regions and small quantities of high-quality training data collected from the target region, an accuracy as high as 88% can be achieved in the classification of marsh types, specifically high marsh and low marsh, at a spatial resolution of 0.6 m. This study contributes to the field of marsh mapping by highlighting the potential of combining multispectral satellite imagery and deep learning for accurate and efficient marsh type classification.

Utilizing Electromagnetic Radiation in Remote Sensing for Vegetation Health Analysis Using NDVI Approach with Sentinel-2 Imagery

The process of utilizing electromagnetic radiation from the radiometric values (Near-Infrared/NIR) and RED in the Normalized Difference Vegetation Index (NDVI) on Sentinel-2 imagery involves measuring light reflection at specific wavelengths reflected by vegetation. NDVI is computed by taking the difference between NIR and RED reflectance, then dividing it by their sum. In-depth analysis of electromagnetic wave interaction and plant spectral characteristics, such as NIR light reflection in healthy leaves, enables mapping vegetation health levels. The research aims to process image data and create a model visualizing vegetation health in the Sentinel-2 research area. Across the total area of 601,971,281.26 m2, 30.37% is healthy vegetation, 12.52% is moderate, and 53.33% is low or unhealthy. The least healthy area has an NDVI of -0.2006, while the healthiest area has a value of 0.6043.

ESTIMATION OF COASTAL WATERS TURBIDITY USING SENTINEL-2 IMAGERY

Turbidity is an important water quality parameter and an indicator of water pollution. Marine remote sensing techniques has become a useful tool for mapping of turbidity at coastal waters. The advantage of using remote sensing for water quality analysis is its ability to obtain synoptic data from the entire study area to produce continuous surface data, can shows detailed spatial variability and periodically. The empirical modeling has been applied in this study to formulate the mathematical relationship between coastal waters turbidity with Sentinel-2 reflectance. This study integrated field survey and image processing. Measurement of in-situ turbidity was done in accordance with imagery acquisition time. Imageries used for this study were Sentinel-2 level-2A. The mathematical relationship was obtained by multiple linear regression model between turbidity and Sentinel-2 reflectance. A mathematical model has been developed in Sentinel-2 imagery and successfully applied to obtain surface turbidity. Estimated turbidity derived from Sentinel-2 imagery is very close to observed turbidity so the proposed model can be used to retrieve turbidity of coastal waters.

Deep learning-based burned forest areas mapping via Sentinel-2 imagery: a comparative study.

In order to evaluate the effects of forest fires on the dynamics of the function and structure of ecosystems, it is necessary to determine burned forest areas with high accuracy, effectively, economically, and practically using satellite images. Extraction of burned forest areas utilizing high-resolution satellite images and image classification algorithms and assessing the successfulness of varied classification algorithms has become a prominent research field. This study aims to indicate on the capability of the deep learning-based Stacked Autoencoders method for the burned forest areas mapping from Sentinel-2 satellite images. The Stacked Autoencoders, used in this study as an unsupervised learning method, were compared qualitatively and quantitatively with frequently used supervised learning algorithms (k-Nearest Neighbors (k-NN), Subspaced k-NN, Support Vector Machines, Random Forest, Bagged Decision Tree, Naive Bayes, Linear Discriminant Analysis) on two distinct burnt forest zones. By selecting burned forest zones with contrasting structural characteristics from one another, an objective assessment was achieved. Manually digitized burned areas from Sentinel-2 satellite images were utilized for accuracy assessment. For comparison, different classification performance and quality metrics (Overall Accuracy, Mean Squared Error, Correlation Coefficient, Structural Similarity Index Measure, Peak Signal-to-Noise Ratio, Universal Image Quality Index, and KAPPA metrics) were used. In addition, whether the Stacked Autoencoders method produces consistent results was examined through boxplots. In terms of both quantitative and qualitative analysis, the Stacked Autoencoders method showed the highest accuracy values.

DEVELOPMENT OF GOOGLE EARTH ENGINE APPLICATION FOR SPATIOTEMPORAL ANALYSIS OF TURBIDITY IN BATAN ESTUARY, AKLAN THROUGH THE HARMONIZATION OF LANDSAT AND SENTINEL-2 IMAGERY

Abstract. Water turbidity poses a threat to marine life and the economy. To address this, regular monitoring is crucial. Field methods are commonly employed, but they can be costly and challenging to perform consistently and historically. Luckily, remote sensing satellites offer a viable solution by collecting image data from space for Earth observation. Researchers have utilized these satellite images to develop indices for analyzing water turbidity and chlorophyll-a content. Leveraging the Google Earth Engine geospatial data computing platform, an application has been created to facilitate water quality monitoring. The app harmonizes Landsat and Sentinel-2 Imagery to improve temporal resolution and increase historical data availability. These images are then subjected to indices, namely the Normalized Difference Turbidity Index (NDTI) and the Normalized Difference Chlorophyll Index (NDCI), to estimate water turbidity and chlorophyll-a content, respectively. To evaluate the effectiveness of the app, the Batan Estuary in Aklan was selected as a test site. Annual mean water turbidity data from 2000 to 2021 were processed and downloaded through the app and subsequently imported into ArcGIS Pro for further analysis. Raster statistics revealed that turbidity and chlorophyll-a content have consistently decreased since 2000, with already low levels observed. Although the river section exhibited the highest turbidity relative to the estuary, it can still be deemed nonpolluted based on US EPA standards. Moreover, ArcGIS's emerging hotspot analysis indicated that the estuary predominantly featured cold spots, indicating minimal areas with clusters of turbid water over time.

State-wide forest canopy height and aboveground biomass map for New York with 10 m resolution, integrating GEDI, Sentinel-1, and Sentinel-2 data

Investigating the quantity of forest aboveground biomass (AGB) is crucial for understanding the role forests play in the global carbon cycle. The canopy height model (CHM) is a critical component in estimating AGB, as it provides a three-dimensional representation of the tree canopy. Traditional CHM estimation methods are time-consuming, labor-intensive, and expensive, particularly at large-scales. Remote sensing is a cost-effective and efficient alternative approach, providing valuable information over large areas in a timely manner. The Global Ecosystem Dynamics Investigation (GEDI) onboard the International Space Station is a space-based light detection and ranging (LiDAR) system designed to collect information on vertical structures of vegetation. One major problem with the collection of GEDI data is that it provides limited information over discrete ground samples, also known as footprints, and thus do not provide wall-to-wall gridded height products. The objective of this study was twofold: a) to integrate the GEDI LiDAR footprint heights with Sentinel-2 multispectral imagery to generate a 10 m wall-to-wall CHM map of New York State (NYS), USA for the year 2019 and b) to improve our previously generated AGB map (both accuracy and resolution) of NYS for the year 2019 by fusing Sentinel-2 multispectral, Sentinel-1 synthetic aperture radar (SAR), and the produced CHM. To generate the 10 m CHM map, the GEDI footprints height measurements were extrapolated using Sentinel-2 imagery and a random forest model. The CHM that was produced was assessed by using GEDI footprints that were not part of the training phase and were therefore independent (extrapolated). Comparing our 10 m CHM with the available global 30 m CHM map provided by Potapov et al. (2021) over NYS shows significant improvement not only in terms of spatial resolution, but also in terms of accuracy. The root mean square error (RMSE) of our 10 m CHM is 4.4 m while this value is 7.49 m for the 30 m CHM over NYS. Similarly, the R2 value for the 10 m CHM map is 0.74, while that of the 30 m CHM is 0.46. Finally, the integration of produced 10 m CHM, Sentinel-1, and Sentinel-2 datasets were utilized to create a 10 m AGB map of NYS with the RMSE of 39.49 Mg/ha, and R2 of 0.65. The results demonstrate the potential of integrating GEDI, Sentinel-1, and Sentinel-2 data for providing a valuable tool for large-scale mapping of forest canopy structure and biomass, which can help to inform forest management and carbon accounting efforts.

Mapping Tree Mortality Caused by Siberian Silkmoth Outbreak Using Sentinel-2 Remote Sensing Data

The Siberian silkmoth is one of the most dangerous coniferous forests pests. Siberian silkmoth outbreaks cause massive defoliation and subsequent forest fires over vast areas. Remote forest disturbance assessments performed after an outbreak make it possible to assess carbon emissions and the potential for natural regeneration, estimate forest fire danger, and reveal the need to implement forest management practices. The goal of the present research was to investigate the use of modern satellite imagery of medium spatial resolution to estimate the percentage of dead trees in a given area. The subject of this study is the Siberian silkmoth outbreak that occurred in 2018–2020 and covered 42 thousand ha in the Irbey region of the Krasnoyarsk Krai. Imagery from the Sentinel-2/MSI sensor was used to calculate a number of spectral indices for images received before and after the outbreak. Field study data were used to create regression models relating the index values to the percentage of dead trees. A number of spectral indices, such as NDVI, dNDVI, NBR, dNBR, NDMI, EVI, and TCG, were used. As a result, spectral indices based on the data from NIR/SWIR bands (NBR, NDMI, dNBR) demonstrated the best correlations with field-measured tree mortality. Therefore, these indices may be used to accurately estimate the percentage of dead trees by remote sensing data. The best was the NBR index with an R2 equal to 0.87, and the lowest RMSE and MAE errors. Consequently, Sentinel-2 imagery can be successfully used for tree mortality assessment over large inaccessible areas disturbed by Siberian silkmoth outbreaks at a relatively low cost.

A deep learning framework to map riverbed sand mining budgets in large tropical deltas

ABSTRACT Rapid urbanization has dramatically increased the demand for river sand, leading to soaring sand extraction rates that often exceed natural replenishment in many rivers globally. However, our understanding of the geomorphic and social-ecological impacts arising from Sand Mining (SM) remains limited, primarily due to insufficient data on sand extraction rates. Conventionally, bathymetry surveys and compilation of declared amounts have been used to quantify SM budgets, but they are often costly and laborious, or result in inaccurate quantification. Here, for the first time, we developed a Remote Sensing (RS)-based Deep Learning (DL) framework to map SM activities and budgets in the Vietnamese Mekong Delta (VMD), a global SM hotspot. We trained a near real-time object detection system to identify three boat classes in Sentinel-1 imagery: Barge with Crane (BC), Sand Transport Boat (STB), and other boats. Our DL model achieved a 96.1% Mean Average Precision (mAP) across all classes and 98.4% for the BC class, used in creating an SM boat density map at an Intersection over Union (IoU) threshold of 0.50. Applying this model to Sentinel-1, 256,647 boats were detected in the VMD between 2014–2022, of which 17.4% were BC. Subsequently, the annual SM budget was estimated by correlating it with a recent riverbed incision map. Our results showed that, between 2015–2022, about 366 Mm3 of sand has been extracted across the VMD. The annual budget has progressively increased from 34.92 Mm3 in 2015 to 53.25 Mm3 in 2022 (by 52%), with an annual increment of around 2.79 Mm3. At the provincial-scale, Dong Thap, An Giang, Vinh Long, Tien Giang, and Can Tho were the locations of intensive mining, accounting for 89.20% of the total extracted volume in the VMD. Finally, our estimated budgets were validated with previous research that yielded a correlation coefficient of 0.99% (with bias of 2.65%). The automatic DL framework developed in this study to quantify SM budgets has a high potential to be applied to other deltas worldwide also facing intensive SM.

Novel oil spill indices for sentinel-2 imagery: A case study of natural seepage in Qaruh Island, Kuwait

Oil spills are a paramount and immediate challenge affecting marine ecosystems globally. Effective and timely monitoring tools, such as oil detection indices, offer a swift means to track oil spill spread across vast oceanic expanses. Moreover, these indices enhance data clarity, making it more conducive for machine learning and deep learning algorithms.This study leverages the natural seepage occurring around Qaruh Island, Kuwait as a unique context for the spectral analysis of oil spills using Sentinel-2 multispectral imagery due to repeated occurrences in the same region. This research evaluated 859 single band and 455 multichannel combinations to identify the most effective combinations in oil-water separability, employing the Jeffries-Matusita (JM) distance measure as a key metric. Bands 1, 2, 3, 8A, 11, and 12 consistently featured among the top-performing indices combinations B1−B11B1+B11;B1+B2B3+B11;B1+B2B3+B12;B1+B2B3+B8A affirming the significant effect of oil spills on visible, Near-Infrared (NIR), and Shortwave Infrared (SWIR) bands. Notably, the indices developed in this study outperformed those from prior research in terms of suitability to unsupervised classification algorithms. A significant conclusion of this study is that incorporating a higher number of bands in the analysis did not correlate with an increase in JM values, suggesting that the selection of specific, informative bands is more critical than the volume of input data. These findings underscore the indispensable role of Sentinel-2 imagery in environmental investigations and highlight the potential for focused, efficient analysis using strategic band combinations for effective oil spill detection.•This study identified optimized Sentinel-2 band combinations for oil-water separability, benefiting from naturally occurring spills around Qaruh Island.•The proposed indices outperformed the previous indices for oil spill visualization and clustering.•The new indices highlighted the critical role of specific band selection over the volume of input data for effective oil spill detection.

Mapping the spatiotemporal patterns of tillage practices across Chinese croplands with Google Earth Engine

ContextTillage practices have major impacts on soil properties, grain yield, soil carbon sequestration, and greenhouse gas emissions. Conservation tillage is an agriculture practice aiming to minimize soil disturbance and protect soil structure. It is reported that more and more croplands are being converted from conventional tillage to conservation tillage in China. However, the information is lacking as to where and when conservation tillage practice is implemented at a regional and national scale. ObjectiveThe objectives of this study are 1) to develop a new framework for classifying tillage practices with a random forest model and Google Earth Engine platform to overcome the limitations in mapping tillage practices across China; and 2) to map the spatiotemporal patterns of tillage practices across Chinese croplands from 2016 to 2020. MethodsWe generated a 1 km grid dataset for tillage practices across Chinese croplands during 2016–2020 based on Sentinel-2 imagery. First, we generated two ground reference sample pools about tillage practices based on literature review and high-resolution maps. Then, we produced Sentinel-2 features during the tillage period by composting 18 bands related to the ratios of returned crop residues. Next, we trained an optimal random forest (RF) classifier by conducting a cross-validation. Finally, we applied the optimal RF classifier to composited features to produce tillage maps from 2016 to 2020. Results and ConclusionsThe validation with ground reference samples shows that the classification results have an overall accuracy of 0.80. Comparisons at a provincial scale show the estimated proportion of conservation tillage is significantly correlated with statistic data with R2 being 0.88. Results show that conservation tillage practice is adopted in about 36% of Chinese croplands, mainly distributed in northeast China, northwest China, and the North China Plain. The proportion of conservation tillage ranged from 29% to 40% during 2016–2020 although with a large spatial heterogeneity. SignificanceThis study develops a sound framework to map tillage practices and provides more details about the spatiotemporal patterns of tillage practices, supporting regional planning of agricultural management and climate change mitigation.

Integration of harvester trajectory and satellite imagery for large-scale winter wheat mapping using deep positive and unlabeled learning

Limited accurate ground truth labels are the primary constraint for data-driven modeling analysis of large-scale crop mapping. Existing labeling methods largely rely on field surveys, visual interpretation, and historical ground information. These labor-intensive approaches are often limited by spatiotemporal heterogeneity of crop distribution and encounter the challenge of gathering extensive crop labels. The massive operating trajectories of agricultural machinery contain precise location information of the crop fields, providing a new source for accurate crop labels at a large spatial scale. This study develops a large-scale crop mapping workflow through widespread harvester trajectory and 10 m Sentinel-2 imagery. The trajectory-based automatic labeling method is developed to generate 287,533 winter wheat labels by jointly using harvester coordinates and satellite images. These generated one-class ground labels are further used to develop positive and unlabeled learning based deep learning models for winter wheat mapping. The Positive and Unlabeled Learning-based Convolutional Neural Network (PUL-CNN) outperforms the other four one-class based classifiers with an F1 score of 94.4 % at 12 study sites. The estimated county-level winter wheat acreage agrees well with census data with R2 of 0.86 in the overall study region. The interpretation analysis based on the Shapley Additive Explanation method shows the heading and greening stages are the critical periods for wheat mapping, aligning well with the separability in Normalized Difference Vegetation Index (NDVI) curves. The results of winter wheat mapping demonstrate the integration of harvester trajectory and remote sensing data facilitates large-scale winter wheat mapping. To the best of our knowledge, this is the first study that fuses operating trajectories of agricultural machinery and satellite images for large-scale crop mapping based on the deep positive and unlabeled learning approach. This study could be possibly applied for better understanding the land cover and land use changes.

Comparison of multi-source satellite remote sensing observations for monitoring the variations of small lakes: a case study of Dai Lai Lake (Vietnam)

Abstract This study compares the capability of Sentinel-1, Sentinel-2, and PlanetScope (PS) satellites in monitoring the variations of surface water of Dai Lai Lake, located in North Vietnam, for the 2018–2023 period. The analysis involves the utilization of Google Earth Engine to partially process Sentinel-1 and Sentinel-2 observations, while PS observations are processed using local computers, to generate VH-polarized backscatter coefficient, Normalized Difference Water Index (NDWI), and Modified of Normalized Difference Water Index (MNDWI) maps. The method for making binary water/non-water maps primarily employs the Otsu algorithm on each single map derived from the previous step. Findings reveal that the lake's water extent remains relatively stable over the 6-year period, and is not strongly affected by the seasonal cycle. Although the spatial distribution patterns of the lake exhibit significant similarity, average water extent of the lake derived from 3-m resolution PS imagery is about 2.17 and 5.60% more than that obtained from 10-m resolution Sentinel-2 and Sentinel-1 imagery, respectively. PS observations are effective for monitoring small lakes, but it is advised to check the quality of its NIR band. Sentinel-2 observations prove great effectiveness for lake monitoring, using both NDWI and MNDWI. For Sentinel-1 observations, potential misclassifications could arise due to similarities in VH-polarized backscatter coefficients between water surfaces and other flat surfaces.