A Dynamic Adaptive Framework for Remote Sensing Imagery Superpixel Segmentation and Classification via Dual-Branch Feature Learning

This article presents an integrated approach for superpixel segmentation (SPS) and classification, leveraging a deep learning (DL) method tailored to high-resolution remote sensing imagery (RSI). The main contributions of this method include designing a SPS approach based on a convolution-based network architecture that directly predicts superpixels on a regular grid, while adding a classification branch that leverages SPS to classify individual superpixels. The proposed method introduces a dynamic adaptive quantization framework and bit mapping modules, enabling the model to flexibly adapt to various bit-width configurations. End-to-end training integrates SPS and classification tasks within the same deep neural network. Comprehensive experiments utilized RSI datasets across three typical scenes: urban, suburban, and agricultural-pastoral areas. Quantitative and qualitative results confirm the superiority for both SPS and semantic segmentation tasks, showing strong potential for scene understanding and land cover classification. Ablation studies further confirm the efficiency and necessity of various components in the model design. This work provides new ideas and technical support for achieving high-precision, fine-grained interpretation of remote sensing scenes.