Yu Wang

Accurate urban land cover information is essential for monitoring urbanization and environmental change, yet the complexity and heterogeneity of urban landscapes remain challenging for high-resolution remote sensing classification. This study used Google Satellite Embeddings as the core feature input, combined with ensemble learning and spatiotemporal post-processing, to generate annual 10 m land cover maps for 37 Chinese metropolises from 2017 to 2024. Results showed that the weighted ensemble model achieved an overall accuracy (OA) of 81.99%, 2.13% higher than the best individual model, while spatiotemporal post-processing further increased OA to 82.59%. Compared with existing land cover products such as FROM-GLC10, the proposed product achieved 4.20% higher OA. The ablation experiment showed that the embedding representation was the primary source of performance improvement. Under the same ensemble learning framework, Satellite Embedding features outperformed conventional remote sensing features by 6.54% in OA. SHapley Additive exPlanations (SHAP) analysis identified A16, A36, A51, A61, and A22 as the key embedding dimensions, while Pearson correlation analysis further revealed associations between the key dimensions and spectral, radar, texture, and topographic information. Change analysis indicated concurrent urban expansion, ecological recovery, and agricultural land contraction from 2017 to 2024, with cropland-to-forest and cropland-to-impervious-surface conversions being the most prominent transition pathways. Overall, Satellite Embeddings provide effective feature representations for annual land cover mapping and long-term change analysis in complex metropolitan environments.

Assessing ecological quality in mining areas is critical for environmental protection and sustainable resource management. However, most previous studies concentrate on large-scale analysis, overlooking fine-scale assessment in mining areas. To address this issue, this study proposed a novel analysis framework for mining areas by integrating high-resolution Landsat data, the Remote Sensing Ecological Index (RSEI), and the Random Forest regression method. Based on the framework, four decades of spatiotemporal dynamics and drivers of ecological quality were revealed in Youjiang River Valley. Results showed that from 1986 to 2024, ecological quality in Youjiang River Valley exhibited a fluctuating upward trend (slope = 0.004/year), with notable improvement concentrated in the most recent decade. Spatially, areas with a significant increasing trend in RSEI (48.71%) were mainly located in natural vegetation regions, whereas areas with a significant decreasing trend (9.11%) were concentrated in impervious surfaces and croplands in northern and central regions. Driver analysis indicates that anthropogenic factors played a crucial role in ecological quality changes. Specifically, land use intensity, precipitation, and sunshine duration were main determinants. These findings offer a comprehensive understanding of ecological quality evolution in subtropical karst mining areas and provide crucial insights for conservation and restoration efforts in Youjiang River Valley.

Timely cropland information is crucial for ensuring food security and promoting sustainable development. Traditional field survey methods are time-consuming and costly, making it difficult to support rapid monitoring of large-scale cropland changes. Furthermore, most existing studies focus on cropland evaluation from a single aspect such as quantity or quality, and thus cannot comprehensively reveal spatiotemporal characteristics of cropland. In this study, a method for evaluating the quantity and quality of cropland using multi-source remote sensing-derived data was proposed and effectively applied in the black soil region in Northeast China. Evaluation results showed that the area of cropland increased significantly in the study area between 2010 and 2018, and the proportion of cropland increased by 1.17%. Simultaneously, cropland patches became larger and landscape connectivity improved. Most of the gained cropland was concentrated in the northeast and west, resulting in a shift in the gravity center of cropland to the northeast direction. Among land converted into cropland, unused land, grassland, and forest were the main sources, accounting for 36.38%, 31.47%, and 16.94% respectively. The quality of cropland in the study area generally improved. The proportion of low-quality cropland decreased by 7.17%, while the proportions of high-quality and medium-quality cropland increased by 5.65% and 5.17%, respectively. Specifically, the quality of cropland improved strongly in the east, improved slightly in the southwest, and declined in the north. Production capacity and soil fertility were key factors impacting cropland quality with obstacle degrees of 36.22% and 15.64%, respectively. Overall, the obtained results were helpful for a comprehensive understanding of spatiotemporal changes in cropland and driving factors and can provide guidance for cropland protection and management. The proposed method demonstrated promising reliability and application potential, which can provide a reference for other cropland evaluation studies.

Accurate information on forest distribution is an essential basis for the protection of forest resources. Recent advances in remote sensing and machine learning have contributed to the monitoring of forest-cover distribution cost-effectively, but reliable methods for rapid forest-cover mapping over mountainous areas are still lacking. In addition, the forest landscape pattern has proven to be closely related to the functioning of forest ecosystems, yet few studies have explicitly measured the forest landscape pattern or revealed its driving forces in mountainous areas. To address these challenges, we developed a framework for forest-cover mapping with multi-source remote sensing data (Sentinel-1, Sentinel-2) and an automated ensemble learning method. We also designed a scheme for forest landscape pattern evaluation and driver attribution based on landscape metrics and random forest regression. Results in the Qilian Mountains showed that the proposed framework and scheme could accurately depict the distribution and pattern of forest cover. The overall accuracy of the obtained level-1 and level-2 forest-cover maps reached 95.49% and 78.05%, respectively. The multi-classifier comparison revealed that for forest classification, the ensemble learning method outperformed base classifiers such as LightGBM, random forests, CatBoost, XGBoost, and neural networks. Integrating multi-dimensional features, including spectral, phenological, topographic, and geographic information, helped distinguish forest cover. Compared with other land-cover products, our mapping results demonstrated high quality and rich spatial details. Furthermore, we found that forest patches in the Qilian Mountains were concentrated in the eastern regions with low-to-medium elevations and shady aspects. We also identified that climate was the critical environmental determent of the forest landscape pattern in the Qilian Mountains. Overall, the proposed framework and scheme have strong application potential for characterizing forest cover and landscape patterns. The mapping and evaluation results can further support forest resource management, ecological assessment, and regional sustainable development.

Irregular spatial dependency is one of the major characteristics of remote sensing images, which brings about challenges for classification tasks. Deep supervised models such as convolutional neural networks (CNNs) have shown great capacity for remote sensing image classification. However, they generally require a huge labeled training set for the fine tuning of a deep neural network. To handle the irregular spatial dependency of remote sensing images and mitigate the conflict between limited labeled samples and training demand, we design a superpixel-guided layer-wise embedding CNN (SLE-CNN) for remote sensing image classification, which can efficiently exploit the information from both labeled and unlabeled samples. With the superpixel-guided sampling strategy for unlabeled samples, we can achieve an automatic determination of the neighborhood covering for a spatial dependency system and thus adapting to real scenes of remote sensing images. In the designed network, two types of loss costs are combined for the training of CNN, i.e., supervised cross entropy and unsupervised reconstruction cost on both labeled and unlabeled samples, respectively. Our experimental results are conducted with three types of remote sensing data, including hyperspectral, multispectral, and synthetic aperture radar (SAR) images. The designed SLE-CNN achieves excellent classification performance in all cases with a limited labeled training set, suggesting its good potential for remote sensing image classification.