Adaptive feature extraction and attention-based segmentation network for remote sensing imagery

Abstract

Semantic segmentation of land cover types is a pivotal task in remote sensing, essential for applications in urban planning, environmental monitoring, disaster management, and agriculture. Accurate segmentation is challenged by imbalanced class distributions and ambiguous boundaries. This paper introduces AdaptiveFusionNet, a novel architecture designed to address heterogeneous complexities in remote sensory image, by leveraging adaptive, multi-scale feature extraction and efficient fusion mechanisms. The architecture comprises three core modules: the Adaptive Pixel Encoder (APE), which enhances pixel-level feature extraction across multiple scales; the Fusion Atrous Pooling (FAP), which effectively integrates contextual information using atrous convolutions; and the Parallel Attention Decoder (PAD), which refines segmentation boundaries through attention-enhanced upsampling. Evaluated on the high-resolution Gaofen 2 dataset, AdaptiveFusionNet demonstrates substantial improvements in key performance metrics, achieving an overall Intersection over Union (IoU) of 71% and excelling in Precision, Recall, and F1 score across various land cover classes, including urban areas, vegetation, water bodies, and infrastructure. An ablation study is presented to validate AdaptiveFusionNet's superiority over existing architectures. The results establish AdaptiveFusionNet as an improved architecture for high-resolution land cover segmentation in terms of both accuracy and computational efficiency.