Very High Resolution Research Articles

The very-high-resolution configuration of the EC-Earth global model for HighResMIP

Abstract. We present the very-high-resolution (VHR) version of the EC-Earth global climate model, EC-Earth3P-VHR, developed for HighResMIP. The model features an atmospheric resolution of ∼16 km and an oceanic resolution of 1/12° (∼8 km), which makes it one of the finest combined resolutions ever used to complete historical and scenario-like CMIP6 simulations. To evaluate the influence of numerical resolution on the simulated climate, EC-Earth3P-VHR is compared with two configurations of the same model at lower resolution: the ∼100 km grid EC-Earth3P-LR (LR) and the ∼25 km grid EC-Earth3P-HR (HR). Of the three configurations, VHR shows the smallest drift in the global mean ocean temperature and salinity at the end of a 100-year 1950s control simulation, which points to a faster equilibrating phase than in LR and HR. In terms of model biases, we compare the historical simulations against observations over the period 1980–2014. In contrast to LR and HR, VHR shows a reduced equatorial Pacific cold tongue bias, an improved Gulf Stream representation with a reduced coastal warm bias and a reduced subpolar North Atlantic cold bias, and more realistic orographic precipitation over mountain ranges. By contrast, VHR shows a larger warm bias and overly low sea ice extent over the Southern Ocean. Such biases in surface temperature have an impact on the atmospheric circulation aloft, connected with a more realistic storm track over the North Atlantic yet a less realistic storm track over the Southern Ocean compared to the lower-resolution model versions. Other biases persist or worsen with increased resolution from LR to VHR, such as the warm bias over the tropical upwelling region and the associated cloud cover underestimation, a precipitation excess over the tropical South Atlantic and North Pacific, and overly thick sea ice and an excess in oceanic mixing in the Arctic. VHR shows improved air–sea coupling over the tropical region, although it tends to overestimate the oceanic influence on the atmospheric variability at midlatitudes compared to observations and LR and HR. Together, these results highlight the potential for improved simulated climate in key regions, such as the Gulf Stream and the Equator, when the atmospheric and oceanic resolutions are finer than 25 km in both the ocean and atmosphere. Thanks to its unprecedented resolution, EC-Earth3P-VHR offers a new opportunity to study climate variability and change of such areas on regional and local spatial scales, in line with regional climate models.

Impact of Downwelling Favorable Winds on Eddy Formation in the West Greenland Current

Abstract Downwelling favorable winds off the west Greenland coast induce onshore Ekman transport, thereby limiting freshwater influx into the interior of the Labrador Sea. Although the eddies shed by West Greenland Current (WGC) provide an important freshwater source, the role of the downwelling favorable winds in eddy formation is still largely unknown. Using the output of a very-high resolution (1/60°) model, combined with idealized model experiments, we find that the downwelling favorable winds push denser water over lighter water, which leads to a gain in Available Potential Energy (APE). Although symmetric instability attenuates eddy formation by flattening isopycnals, the net effect of downwelling favorable wind is a more baroclinic front along the WGC and increased buoyant eddy formation, which provides an additional freshwater flux into the convective interior of the Labrador Sea.

Monitoring and Prediction of Land Surface Phenology Using Satellite Earth Observations—A Brief Review

Monitoring and predicting land surface phenology (LSP) are essential for understanding ecosystem dynamics, climate change impacts, and forest and agricultural productivity. Satellite Earth observation (EO) missions have played a crucial role in the advancement of LSP research, enabling global and continuous monitoring of vegetation cycles. This review provides a brief overview of key EO satellite missions, including the advanced very-high resolution radiometer (AVHRR), moderate resolution imaging spectroradiometer (MODIS), and the Landsat program, which have played an important role in capturing LSP dynamics at various spatial and temporal scales. Recent advancements in machine learning techniques have further enhanced LSP prediction capabilities, offering promising approaches for short-term prediction of vegetation phenology and cropland suitability assessment. Data cubes, which organize multidimensional EO data, provide an innovative framework for enhancing LSP analyses by integrating diverse data sources and simplifying data access and processing. This brief review highlights the potential of satellite-based monitoring, machine learning models, and data cube infrastructure for advancing LSP research and provides insights into current trends, challenges, and future directions.

Monitoring Indian ungauged small reservoirs volume from remote sensing: Feasibility, bias and perspectives

What remote sensing products can be used to better quantify the water stored in hundreds of thousands Indian Small Reservoirs (SR)? This ungauged hydrological component of the water cycle is intermittently filled with rainwater runoff, constantly reshaped by farmers since last two decades, crucial for upstream irrigated agriculture. Given the small size and shallow depth of those reservoirs, usual remote sensing techniques (Altimeters and LIDAR) used in spatial hydrology to monitor their water level are not adapted. We evaluated the uncertainty of SR volume retrieval methods based on surface water estimates from Sentinel-2 and associated volumes from global available DEM at a medium to coarse resolution. Four pair of stereoscopic images at Very High Resolution (VHR) from Pléiades satellites were acquired during the last two dry hydrological years (2016 and 2019), when SR were totally empty. The Pléiades DEMs produced were cross validated with LIDAR IceSAT-2 products, and used to extract 504 SR bathymetries within an area covering 1,813 km2 located in the Telangana state (114,789 km2). We compared Pléiades based retrievals to freely available regional to global DEM to explore the regional volume retrieval Bias: ALOS World 3D-30 m, WorldDEM GLO-30 at 30 m TanDEM-X DEM at 90 m and one Indian DEM (CartoDEM at 30 m). The Mean Absolute Percentage Error (MAPE) of reservoir volumes from global DEMs range from 47% to 78%. MAPE are 17%, 29% and 46% for Pléiades DEM resampled at 12, 30 and 90 m, respectively. In a near future, upcoming stereoscopy satellite missions at lower costs and with larger coverage and shorter revisit such as CO3D will provide 12m or higher resolution DEMs that, if acquired in dry years, will lead to acceptable MAPE (< 20%), to monitor empty SR geometries throughout India and other semi-arid areas in the world.

Contextual Learning in Fourier Complex Field for VHR Remote Sensing Images.

Very high-resolution (VHR) remote sensing (RS) image classification is the fundamental task for RS image analysis and understanding. Recently, Transformer-based models demonstrated outstanding potential for learning high-order contextual relationships from natural images with general resolution ( ≈ 224 × 224 pixels) and achieved remarkable results on general image classification tasks. However, the complexity of the naive Transformer grows quadratically with the increase in image size, which prevents Transformer-based models from VHR RS image ( ≥ 500 × 500 pixels) classification and other computationally expensive downstream tasks. To this end, we propose to decompose the expensive self-attention (SA) into real and imaginary parts via discrete Fourier transform (DFT) and, therefore, propose an efficient complex SA (CSA) mechanism. Benefiting from the conjugated symmetric property of DFT, CSA is capable to model the high-order contextual information with less than half computations of naive SA. To overcome the gradient explosion in Fourier complex field, we replace the Softmax function with the carefully designed Logmax function to normalize the attention map of CSA and stabilize the gradient propagation. By stacking various layers of CSA blocks, we propose the Fourier complex Transformer (FCT) model to learn global contextual information from VHR aerial images following the hierarchical manners. Universal experiments conducted on commonly used RS classification datasets demonstrate the effectiveness and efficiency of FCT, especially on VHR RS images. The source code of FCT will be available at https://github.com/Gao-xiyuan/FCT.

Learning to Holistically Detect Bridges From Large-Size VHR Remote Sensing Imagery.

Bridge detection in remote sensing images (RSIs) plays a crucial role in various applications, but it poses unique challenges compared to the detection of other objects. In RSIs, bridges exhibit considerable variations in terms of their spatial scales and aspect ratios. Therefore, to ensure the visibility and integrity of bridges, it is essential to perform holistic bridge detection in large-size very-high-resolution (VHR) RSIs. However, the lack of datasets with large-size VHR RSIs limits the deep learning algorithms' performance on bridge detection. Due to the limitation of GPU memory in tackling large-size images, deep learning-based object detection methods commonly adopt the cropping strategy, which inevitably results in label fragmentation and discontinuous prediction. To ameliorate the scarcity of datasets, this paper proposes a large-scale dataset named GLH-Bridge comprising 6,000 VHR RSIs sampled from diverse geographic locations across the globe. These images encompass a wide range of sizes, varying from 2,048 × 2,048 to 16,384 × 16,384 pixels, and collectively feature 59,737 bridges. These bridges span diverse backgrounds, and each of them has been manually annotated, using both an oriented bounding box (OBB) and a horizontal bounding box (HBB). Furthermore, we present an efficient network for holistic bridge detection (HBD-Net) in large-size RSIs. The HBD-Net presents a separate detector-based feature fusion (SDFF) architecture and is optimized via a shape-sensitive sample re-weighting (SSRW) strategy. The SDFF architecture performs inter-layer feature fusion (IFF) to incorporate multi-scale context in the dynamic image pyramid (DIP) of the large-size image, and the SSRW strategy is employed to ensure an equitable balance in the regression weight of bridges with various aspect ratios. Based on the proposed GLH-Bridge dataset, we establish a bridge detection benchmark including the OBB and HBB tasks, and validate the effectiveness of the proposed HBD-Net. Additionally, cross-dataset generalization experiments on two publicly available datasets illustrate the strong generalization capability of the GLH-Bridge dataset.

A semi-automated method using object-based image analysis (OBIA) to detect and enumerate beluga whales in summer from very high-resolution (VHR) satellite imagery.

Very high-resolution (VHR) satellite imagery has proven to be useful for detection of large to medium cetaceans, such as odontocetes and offers some significant advantages over traditional detection methods. However, the significant time investment needed to manually read satellite imagery is currently a limiting factor to use this method across large open ocean regions. The objective of this study is to develop a semi-automated detection method using object-based image analysis to identify beluga whales (Delphinapterus leucas) in open water (summer) ocean conditions in the Arctic using panchromatic WorldView-3 satellite imagery and compare the detection time between human read and algorithm detected imagery. The false negative rate, false positive rate, and automated count deviation were used to assess the accuracy and reliability of various algorithms for reading training and test imagery. The best algorithm, which used spectral mean and texture variance attributes, detected no false positives and the false negative rate was low (<4%). This algorithm was able to accurately and reliably identify all the whales detected by experienced readers in the ice-free panchromatic image. The autodetection algorithm does have difficulty separately identifying whales that are perpendicular to one another, whales below the surface, and may use multiple segments to define a whale. As a result, for determining counts of whales, a reader should manually review the automated results. However, object-based image analysis offers a viable solution for processing large amounts of satellite imagery for detecting medium-sized beluga whales while eliminating all areas of the imagery which are whale-free. This algorithm could be adapted for detecting other cetaceans in ice-free water.

Exploring the Relationship Between Very-High-Resolution Satellite Imagery Data and Fruit Count for Predicting Mango Yield at Multiple Scales

Tree- and block-level prediction of mango yield is important for farm operations, but current manual methods are inefficient. Previous research has identified the accuracies of mango yield forecasting using very-high-resolution (VHR) satellite imagery and an ’18-tree’ stratified sampling method. However, this approach still requires infield sampling to calibrate canopy reflectance and the derived block-level algorithms are unable to translate to other orchards due to the influences of abiotic and biotic conditions. To better appreciate these influences, individual tree yields and corresponding canopy reflectance properties were collected from 2015 to 2021 for 1958 individual mango trees from 55 orchard blocks across 14 farms located in three mango growing regions of Australia. A linear regression analysis of the block-level data revealed the non-existence of a universal relationship between the 24 vegetation indices (VIs) derived from VHR satellite data and fruit count per tree, an outcome likely due to the influence of location, season, management and cultivar. The tree-level fruit count predicted using a random forest (RF) model trained on all calibration data produced a percentage root mean squared error (PRMSE) of 26.5% and a mean absolute error (MAE) of 48 fruits/tree. The lowest PRMSEs produced from RF-based models developed from location, season and cultivar subsets at the individual tree level ranged from 19.3% to 32.6%. At the block level, the PRMSE for the combined model was 10.1% and the lowest values for the location, seasonal and cultivar subset models varied between 7.2% and 10.0% upon validation. Generally, the block-level predictions outperformed the individual tree-level models. Maps were produced to provide mango growers with a visual representation of yield variability across orchards. This enables better identification and management of the influence of abiotic and biotic constraints on production. Future research could investigate the causes of spatial yield variability in mango orchards.

Land covers associated with forest expansion hot spots in the Nepal Himalaya

Abstract Many regions of the Himalaya are experiencing forest expansion, signifying a widespread forest transition across this fragile mountain system. In the Nepal Himalaya, forest expansion is increasingly attributed to the narrative of cropland abandonment driven by rural outmigration. Understanding forest change dynamics as consequences of land use changes is critical in the context of rural mountain societies such as Nepal, which exhibit complex interactions between forest, agriculture, and livelihoods. We analysed forest gain in Nepal from 2000–2019 to inquire where and when forests have expanded and what land covers transitioned into forests. We integrated spatiotemporal analysis of national land cover maps and photointerpretation of very high resolution (VHR) imagery to improve analytical estimates. The Middle Mountain physiographic region experienced the strongest hot spots of forest gains relative to the high mountains or lowlands. Visual interpretation of VHR showed that 68.8% of forest gains since 2000 occurred on shrubland, and 26.5% on cropland. The dominance of shrubland-forest transitions implied that shrubland is an intermediate stage in a multi-decadal succession process, or a semi-permanent state such as arrested succession. Notably however, we also detected cropland-forest transitions occurring as fast as 6–10 years, indicating the potential for rapid forest succession if biophysical conditions permit. Thus, variations in land cover transitions to forests in the hills implicated multiple pathways of forest regeneration, reflecting a variety of spatiotemporally explicit drivers. Our results differ from previous studies because VHR image analysis accounted for land cover classification errors and improved land cover quantification. Further research on spatially explicit mechanisms and drivers of forest gain are needed to understand the synergies of forest, agriculture, and livelihoods to inform land use land cover policies that could be leveraged to enhance rural mountain livelihoods.

Counting Cattle in Argentina with AI and Satellite Images

Abstract. From a specific need of the national institution responsible for plant and animal health in Argentina, SENASA, this collaborative work between the public sectors together with academia arises. The objective was the development of a Deep Learning algorithm for the detection of livestock, cows in particular, in very high-resolution satellite images (provided by the Argentine Space Agency (CONAE)) and its subsequent counting. The basic elements involved in the development of artificial intelligence are detailed, such as the selection and acquisition of satellite images, their very thorough preprocessing and labelling, details of the training stage and some forms of error quantification. The image database is made up of about 320 scenes (based on very high resolution (VHR) satellite data from the Pleiades and Pleiades NEO sensors, ©Airbus 2022, distributed by CONAE) of the Pampas and Patagonian regions in Argentina. Around 8,000 labels of different types of animals were generated, the most common were cows and sheep. Also labels that did not represent animals to contribute to training. Finally, some very promising preliminary results are presented, such as the average error in the count that was achieved was ± four cows. The usefulness of these tools is reflected in better management of renewable natural resources linked to animal health issues, fiscal issues and within the framework of the 2030 Agenda and Sustainable Development Goals #12, #15 and #17.

Precise City-Scale Urban Water Body Semantic Segmentation and Open-Source Sampleset Construction Based on Very High-Resolution Remote Sensing: A Case Study in Chengdu

Addressing the challenges related to urban water bodies is essential for advancing urban planning and development. Therefore, obtaining precise and timely information regarding urban water bodies is of paramount importance. To address issues such as incomplete extraction boundaries, mistaken feature identification, and omission of small water bodies, this study utilized very high-resolution (VHR) satellite images of the Chengdu urban area and its surroundings to create the Chengdu Urban Water Bodies Semantic Segmentation Dataset (CDUWD). Based on the shape characteristics of water bodies, these images were processed through annotation, cropping, and other operations. We introduced Ad-SegFormer, an enhanced model based on SegFormer, which integrates a densely connected atrous spatial pyramid pooling module (DenseASPP) and progressive feature pyramid network (AFPN) to better handle the multi-scale characteristics of urban water bodies. The experimental results demonstrate the effectiveness of combining the CDUWD dataset with the Ad-SegFormer model for large-scale urban water body extraction, achieving accuracy rates exceeding 96%. This study demonstrates the effectiveness of Ad-SegFormer in improving water body extraction and provides a valuable reference for extracting large-scale urban water body information using VHR images.

Cloud Detection Using a UNet3+ Model with a Hybrid Swin Transformer and EfficientNet (UNet3+STE) for Very-High-Resolution Satellite Imagery

It is necessary to extract and recognize the cloud regions presented in imagery to generate satellite imagery as analysis-ready data (ARD). In this manuscript, we proposed a new deep learning model to detect cloud areas in very-high-resolution (VHR) satellite imagery by fusing two deep learning architectures. The proposed UNet3+ model with a hybrid Swin Transformer and EfficientNet (UNet3+STE) was based on the structure of UNet3+, with the encoder sequentially combining EfficientNet based on mobile inverted bottleneck convolution (MBConv) and the Swin Transformer. By sequentially utilizing convolutional neural networks (CNNs) and transformer layers, the proposed algorithm aimed to extract the local and global information of cloud regions effectively. In addition, the decoder used MBConv to restore the spatial information of the feature map extracted by the encoder and adopted the deep supervision strategy of UNet3+ to enhance the model’s performance. The proposed model was trained using the open dataset derived from KOMPSAT-3 and 3A satellite imagery and conducted a comparative evaluation with the state-of-the-art (SOTA) methods on fourteen test datasets at the product level. The experimental results confirmed that the proposed UNet3+STE model outperformed the SOTA methods and demonstrated the most stable precision, recall, and F1 score values with fewer parameters and lower complexity.

An Object-Aware Network Embedding Deep Superpixel for Semantic Segmentation of Remote Sensing Images

Semantic segmentation forms the foundation for understanding very high resolution (VHR) remote sensing images, with extensive demand and practical application value. The convolutional neural networks (CNNs), known for their prowess in hierarchical feature representation, have dominated the field of semantic image segmentation. Recently, hierarchical vision transformers such as Swin have also shown excellent performance for semantic segmentation tasks. However, the hierarchical structure enlarges the receptive field to accumulate features and inevitably leads to the blurring of object boundaries. We introduce a novel object-aware network, Embedding deep SuperPixel, for VHR image semantic segmentation called ESPNet, which integrates advanced ConvNeXt and the learnable superpixel algorithm. Specifically, the developed task-oriented superpixel generation module can refine the results of the semantic segmentation branch by preserving object boundaries. This study reveals the capability of utilizing deep convolutional neural networks to accomplish both superpixel generation and semantic segmentation of VHR images within an integrated end-to-end framework. The proposed method achieved mIoU scores of 84.32, 90.13, and 55.73 on the Vaihingen, Potsdam, and LoveDA datasets, respectively. These results indicate that our model surpasses the current advanced methods, thus demonstrating the effectiveness of the proposed scheme.

Feature-Selection-Based Unsupervised Transfer Learning for Change Detection from VHR Optical Images

Accurate understanding of urban land use change information is of great significance for urban planning, urban monitoring, and disaster assessment. The use of Very-High-Resolution (VHR) remote sensing images for change detection on urban land features has gradually become mainstream. However, most existing transfer learning-based change detection models compute multiple deep image features, leading to feature redundancy. Therefore, we propose a Transfer Learning Change Detection Model Based on Change Feature Selection (TL-FS). The proposed method involves using a pretrained transfer learning model framework to compute deep features from multitemporal remote sensing images. A change feature selection algorithm is then designed to filter relevant change information. Subsequently, these change features are combined into a vector. The Change Vector Analysis (CVA) is employed to calculate the magnitude of change in the vector. Finally, the Fuzzy C-Means (FCM) classification is utilized to obtain binary change detection results. In this study, we selected four VHR optical image datasets from Beijing-2 for the experiment. Compared with the Change Vector Analysis and Spectral Gradient Difference, the TL-FS method had maximum increases of 26.41% in the F1-score, 38.04% in precision, 29.88% in recall, and 26.15% in the overall accuracy. The results of the ablation experiments also indicate that TL-FS could provide clearer texture and shape detections for dual-temporal VHR image changes. It can effectively detect complex features in urban scenes.

REWARE: a seismic processing algorithm to retrieve geological information from the water column

SUMMARY When interpreting marine very high-resolution (VHR) single-channel seismic reflection data, the signal in the water column is generally considered as noise and is often eliminated by a water-mute application to focus on geological information under the seafloor. Alternatively, the signal in the water column can be used to study ocean currents or gas/fluid emissions. To provide images of the sedimentary formations and tectonic structures beneath the seafloor in shallow water regions, such as continental shelves and lakes, marine seismic reflection profiles are often acquired using a single-channel streamer and sparker-type source, providing VHR data, with limited penetration depth. To exploit the full potential of these single-channel data, we propose a simple algorithm, called REWARE (Recovery of Water-column Acoustic Reflectors). This algorithm allows to extract further geological information from the water-column data using open-source codes (Seismic Un*x), adding the coherent signal from the previous shots, recorded in the water column, to the previous traces. The record length becomes longer while maintaining a very high trace-to-trace consistency. To demonstrate its efficiency, we present two examples of the REWARE processing in two different geological contexts: the East Sardinia shelf (Italy) and the North Evia Gulf (Greece). This method provides deeper images than with original data for seismic data acquired across steep slopes, such as canyons or continental shelf breaks. Thus, depending on the seafloor geometry and subseafloor structures, it is possible to image or map sediment layers and tectonic structures at depth, keeping a very high structural resolution.

Enhancing Pléiades-based crop mapping with multi-temporal and texture information

Accurate crop mapping using satellite imagery is crucial for improving the monitoring of agricultural landscapes. Very high resolution (VHR) satellite imagery offers unique capabilities in this respect, allowing for even small fields to be discerned and image texture analysis to be performed. Additionally, satellite imagery has greater efficiency than unmanned aerial vehicles due to its extensive coverage. Moreover, the operation flexibility of VHR satellites means that timely image acquisition is possible several times during the growing season. This study investigates the potential of VHR Pléiades images and the random forest classifier for accurate crop mapping. Four images acquired on April 9th, May 12th, May 31st, and June 20th were used to test 16 classification scenarios, including single-date and multi-temporal combinations of spectral bands, texture features, and vegetation Indices (VIs). The classification using the spectral bands from all four images achieved the highest overall accuracy, 93.9% and 96.3% at field and pixel levels, respectively. The bitemporal classifications had lower accuracy. Nevertheless, the combination of the May 12th and June 20th spectral bands had 90% accuracy, which indicated that two images may be sufficient for reliable mapping if the periods with phenological differences between crops are considered. Adding texture features to the spectral bands significantly enhanced the accuracy (up to 8%) of single-date classifications, making it highly recommended when only one image is available. However, the impact of texture was more pronounced on the later dates. It showed the most marked benefit for vineyards and alfalfa, with minimal or no improvement observed for other classes like winter barley. An additional increase in overall accuracy was achieved in three of the four dates by supplementing the spectral and texture bands with VIs. This study highlights the importance of considering image acquisition dates and crop types when designing satellite-based crop mapping strategies for optimal accuracy.

Cropland inundation mapping in a mountain dominated region based on multi-resolution remotely sensed imagery and active learning for semantic segmentation

Accurate and timely mapping of cropland inundated area is critical to loss assessment and recovery plan development. However, existing methods for cropland inundation mapping using Synthetic Aperture Radar (SAR) imagery are primarily designed for plain regions and may not be suitable for mountainous regions. The complex rugged terrain at large scale present challenges such as significant noise from radar shadows, smooth surfaces, floodplains, and permanent water bodies. To remove these noise for effectively mapping of cropland inundation, this study proposes a robust two-stage framework integrating the multi-resolution remotely sensed imageries and semantic segmentation method. First, active learning for cropland extraction is performed with Google Very-High-Resolution (VHR) imagery and Segformer model. Second, the cropland inundation area is obtained using marker-controlled watershed segmentation (MCW) on Sentinel-1 SAR imagery with area-of-interest (AOI) filtering. Results show that the extracted cropland mask are superior to six public cropland data products including ESA, ESRI, FCS30, FROM, GL30, and SLOPE. In addition, with AOI filtering, the F-score of the potential flood identification using MCW was increased from 0.685 to 0.828. At last, the monthly cropland inundation caused by intense rainfall varied approximately from 6.6 km2 to 436.1 km2 during monsoon season of 2017–2022 in Guangdong. Our approach offers a fast and feasible solution for mapping cropland inundation in large and complex mountainous regions, without requiring extensive parameter tuning or additional data.

Global road extraction using a pseudo-label guided framework: from benchmark dataset to cross-region semi-supervised learning

ABSTRACT Recent advancements in satellite remote sensing technology and computer vision have enabled rapid extraction of road networks from massive, Very High-Resolution (VHR) satellite imagery. However, current road extraction methods face the following limitations: 1) Insufficient availability of accurate and diverse training datasets for global-scale road extraction; 2) Costly and time-consuming manual labeling of millions of road samples; and 3) Limited generalization ability of deep learning models across diverse global contexts, resulting in better performance for regions well-represented in the training dataset, but worse performance when faced with domain gaps. To address these challenges, a semi-supervised framework was developed in this study, which includes a global-scale benchmark dataset – termed GlobalRoadSet (GRSet) – and a pseudo-label guided semi-supervised road extraction network – termed GlobalRoadNetSF (GRNetSF). The GRSet dataset was constructed using high-resolution satellite imagery and open-source crowdsourced OpenStreetMap (OSM) data. It comprises 47,210 samples collected from 121 capital cities across six populated continents. The GRNetSF trains the network by generating pseudo-labels for unlabeled images, combined with a few labeled samples from the target region. To enhance the quality of the pseudo-labels, strong data augmentation perturbation and auxiliary feature perturbation techniques are employed to ensure model prediction consistency. The proposed GRNetSF_GRSet framework was implemented in over 30 cities worldwide, where most of the Intersection-over-Union (IoU) values increased by more than 10%. This outcome confirms its strong generalization ability.

Uncovering mangrove range limits using very high resolution satellite imagery to detect fine‐scale mangrove and saltmarsh habitats in dynamic coastal ecotones

AbstractMangroves are important ecosystems for coastal biodiversity, resilience and carbon dynamics that are being threatened globally by human pressures and the impacts of climate change. Yet, at several geographic range limits in tropical–temperate transition zones, mangrove ecosystems are expanding poleward in response to changing macroclimatic drivers. Mangroves near range limits often grow to smaller statures and form dynamic, patchy distributions with other coastal habitats, which are difficult to map using moderate‐resolution (30‐m) satellite imagery. As a result, many of these mangrove areas are missing in global distribution maps. To better map small, scrub mangroves, we tested Landsat (30‐m) and Sentinel (10‐m) against very high resolution (VHR) Planet (3‐m) and WorldView (1.8‐m) imagery and assessed the accuracy of machine learning classification approaches in discerning current (2022) mangrove and saltmarsh from other coastal habitats in a rapidly changing ecotone along the east coast of Florida, USA. Our aim is to (1) quantify the mappable differences in landscape composition and complexity, class dominance and spatial properties of mangrove and saltmarsh patches due to image resolution; and (2) to resolve mapping uncertainties in the region. We found that the ability of Landsat to map mangrove distributions at the leading range edge was hampered by the size and extent of mangrove stands being too small for detection (50% accuracy). WorldView was the most successful in discerning mangroves from other wetland habitats (84% accuracy), closely followed by Planet (82%) and Sentinel (81%). With WorldView, we detected 800 ha of mangroves within the Florida range‐limit study area, 35% more mangroves than were detected with Planet, 114% more than Sentinel and 537% more than Landsat. Higher‐resolution imagery helped reveal additional variability in landscape metrics quantifying diversity, spatial configuration and connectedness among mangrove and saltmarsh habitats at the landscape, class and patch scales. Overall, VHR satellite imagery improved our ability to map mangroves at range limits and can help supplement moderate‐resolution global distributions and outdated regional maps.

A Scene Unmixing Deep Learning Network for Local Climate Zone Mapping and Analysis Using Very High Resolution Remote Sensing Imagery

Abstract. With the acceleration of urbanization, the environment and climate necessary for our survival have gradually deteriorated, leading to the increasing prominence of the Urban Heat Island (UHI) effect. Local Climate Zone (LCZ) classification, as a standard of urban morphology, has become an essential tool for monitoring the UHI effect and conducting temperature studies. Deep Learning (DL) models have the ability to represent high-level semantic features. Therefore, this paper proposes amixed scene unmixing DL framework for LCZ mapping and analysis using Very High Resolution (VHR) remote sensing images. This framework consists of a two-stream deep network, including a pure scene classification network (PS-Net) and a mixed scene unmixing network (MSU-Net). We conducted random sampling tests in Wuhan, China in the experiment A. The results show that this model achieved a satisfactory accuracy with the Overall Accuracies (OAs) is 96.78% and a mixed scene unmixing Mean Absolute Error (MAE) of 0.0495. Furthermore, we applied the proposed model to generate LCZ map for five districts in Wuhan in the experiment B. The test accuracy between two experiments differs very slightly. These results demonstrate the applicability and potential of our model for LCZ mapping and urban climate analysis.