Deep convolutional neural networks have promoted significant progress in building extraction from high-resolution remote sensing imagery. Although most of such work focuses on modifying existing image segmentation networks in computer vision, we propose a new network in this paper, Deep Encoding Network (DE-Net), that is designed for the very problem based on many lately introduced techniques in image segmentation. Four modules are used to construct DE-Net: the inception-style downsampling modules combining a striding convolution layer and a max-pooling layer, the encoding modules comprising six linear residual blocks with a scaled exponential linear unit (SELU) activation function, the compressing modules reducing the feature channels, and a densely upsampling module that enables the network to encode spatial information inside feature maps. Thus, DE-Net achieves state-of-the-art performance on the WHU Building Dataset in recall, F1-Score, and intersection over union (IoU) metrics without pre-training. It also outperformed several segmentation networks in our self-built Suzhou Satellite Building Dataset. The experimental results validate the effectiveness of DE-Net on building extraction from aerial imagery and satellite imagery. It also suggests that given enough training data, designing and training a network from scratch may excel fine-tuning models pre-trained on datasets unrelated to building extraction.

Remote sensing satellites provide observations of the earth’s surface, which are crucial data for applications and analyses in several fields, including agriculture, environmental protection, and sustainable development. However, the wide and frequent occurrence of clouds highly undermines the quality and availability of usable optical data, particularly low-temporal-resolution data. Although deep learning techniques have facilitated recent progress in cloud removal algorithms, thick cloud removal under changing land cover remains challenging. In this study, we propose a framework to remove thick clouds, thin clouds, and cloud shadow from Sentinel-2 images. The framework integrates the spatial detail in a Sentinel-2 reference image and the coarse spectral pattern in a near-target-date Sentinel-3 image as spatiotemporal guidance to generate missing data with land cover change information in a cloudy Sentinel-2 image. The reconstruction is performed using a spatiotemporal attention network (STAN) that adopts the self-attention mechanism, residual learning, and high-pass features to enhance feature extraction from the multisource data. The experimental results show that STAN outperforms ResUnet, CRN, STS-CNN, and DSen2-CR in terms of multiple quantitative metrics and visual characteristics. The comparative experiment proves the integration of Sentinel-3 data improves the cloud removal performance, especially in areas with distinctive and heterogeneous land cover changes under large-scale cloud cover. The experimental results also indicate high generalizability of STAN when the Sentinel-3 image is far from the target date, when transferring features to cloud removal for new images, and even with limited training data that simulates severe cloud cover.

Under increasingly low urban land resources, carrying out roof greening to exploit new green space is a good strategy for sustainable development. Therefore, it is necessary to evaluate the suitability of roof greening for buildings in cities. However, most current evaluation methods are based on qualitative and conceptual research. In this paper, a methodological framework for roof greening suitability evaluation is proposed based on the basic units of building roofs extracted via deep learning technologies. The building, environmental and social criteria related to roof greening are extracted using technologies such as deep learning, machine learning, remote sensing (RS) methods and geographic information system (GIS) methods. The technique for order preference by similarity to an ideal solution (TOPSIS) method is applied to quantify the suitability of each roof, and Sobol sensitivity analysis of the score results is conducted. The experiment on Xiamen Island shows that the final evaluation results are highly sensitive to the changes in weight of the green space distance, population density and the air pollution level. This framework is helpful for the quantitative and objective development of roof greening suitability evaluation.

Deep learning-based road extraction technologies can swiftly identify road information in complex environments, playing a crucial role in advancing urban intelligence and achieving sustainable development goals. However, the current road extraction models exhibit significant omissions when confronted with scenarios involving occluded roads and densely distributed road networks. To address the issue of partial road extraction, this paper introduces a novel method named SSCNet. This approach addresses potential causes of road extraction failures by leveraging synchronous stratification learning and cross-level feature fusion to fully utilize information across various layers. The model incorporates a Dynamic Road Detail Matcher to extract a wealth of road detail information from shallow feature maps, a Cross Contextual Adaptive Attention to capture multi-scale contextual information and interact with shallow feature maps, and a Multi-Scale Global Information Integrator to consolidate global information from different levels of feature maps, enhancing the model’s understanding of road integrity. The model has been extensively tested on the public datasets DeepGlobe, Massachusetts, SpaceNet, and RoadTracer, showing significant improvements in F1 score and IoU compared to current state-of-the-art models.

The precise mapping of crop spatial distribution using remote sensing datasets is a fundamental task in precision agriculture, yet conventional methods based on pixels or segmented objects often neglect the boundary constraints from the cultivated areas and correlations between analysis units. In response, the proposed work is a novel attempt to address parcelwise crop classification with multi-level consistency constraints, generating robust features and consequently more precise results. Pixel-level constraints enforce spectral homogeneity within parcels extracted by the neural network during feature construction. Each parcel is then correlated with its closest neighbors considering comprehensively the spatial, environmental, and temporal similarities, providing consistency information at the parcel scale. Two model structures based primarily on the graph neural network and attention mechanism were proposed for crop identification. Results show that both constraints may yield significant benefits given their enhancement to feature stability and completeness, especially in small-sample cases. The neighborhood models achieved average accuracy improvements of 3.10%, 3.47%, 4.65%, 4.34%, and 15.78%, respectively, under varying training set proportions of 0.5, 0.4, 0.3, 0.2, and 0.1, and an OA of 89.50% could still be maintained for the graph-based one even when the training set ratio was limited to 0.02. Within the calculation, various intelligent algorithms such as CNN, RF, GNN, and attention mechanisms are methodically leveraged for effective pattern recognition of both spatial, sequential, and neighborhood dimensions, achieving an efficacious uncertainty reduction. Overall, the proposed algorithm may serve as a universal framework for parcel-wise type inference, facilitating the effective implementation of precision agriculture.