An Enhanced DeepLabV3 Model with Attention Mechanism and Hierarchical Feature Fusion for Semantic Segmentation of High-Resolution Remote Sensing Imagery+

With the advancement of satellite remote sensing and related techniques, the resolution of remote sensing satellites and aerial remote sensing has been continuously improved, providing more detailed ground observation information. The significance of semantic segmentation for high-resolution remote sensing images has also increased with the expansion of their applications. Hence, a new research direction in image interpretation and remote sensing data processing is to establish a connection between the pixel-level features and higher-level semantics in order to achieve precise semantic segmentation of remote sensing images. It is still difficult to extract fine details from the complicated remote sensing data by applying the traditional region-based and edge-based approaches. Although deep learning can enhance the semantic representation, further improvement in segmentation accuracy is usually accompanied by a larger model size and a heavier computational burden. To overcome this problem, this paper investigates targeted structural modifications of a standard segmentation network, and a specially designed DeepLabV3+ architecture is suggested for high-resolution remote sensing. The SE-ASPP module takes the place of the conventional method to deal with scale differences, while the CFF module integrates the hierarchical features. CARAFE upsampling, which is intended to retain the fine texture features, is utilized to solve the issue of spatial reconstruction. Moreover, an integrated loss function combining Dice and cross-entropy is employed to solve the inherent class imbalance in urban objects. In the end, the improved architecture demonstrates good potential for accurate and efficient remote sensing applications owing to its ability to handle the complex spatial relationships more effectively than the basic one. The effectiveness of this approach is evaluated on the ISPRS Potsdam and Vaihingen datasets using the official ISPRS train/test tile partition, with the proposed model and the DeepLabV3+ baseline averaged over three independent runs and reported as mean ± standard deviation. The improved network attains overall accuracies of 87.98% and 85.58%, mean intersection over union (mIoU) of 73.58% ± 0.03 and 64.99% ± 0.26, and mean F1-scores (mF1) of 83.49% ± 0.03 and 77.21% ± 0.30 on Potsdam and Vaihingen, respectively. Relative to the standard DeepLabV3+ baseline trained under the same protocol, the proposed model improves mIoU by 1.02 and 2.82 points and yields the highest car-class IoU among the four models with per-class results on Potsdam, while its boundary F1-score rises consistently, indicating that the main benefit lies in boundary delineation and small-object recovery rather than in a uniform accuracy lead over all competing models. The outcomes show that the proposed structure provides a balanced and reproducible solution for complex urban scene interpretation, with a moderate and explicitly quantified increase in parameters and computational cost.