RFE-Net: A Lightweight Improved DeepLabV3+ Network with Transformer-Enhanced ASPP for Remote Sensing Rice Field Segmentation

Abstract

Existing semantic segmentation models for rice remote sensing often suffer from inaccurate extraction of highresolution rice fields, large model sizes, and high computational costs. To address these issues, this study proposes a lightweight Rice Field Extraction Network (RFE-Net) based on the DeepLabV3+ framework. The proposed network incorporates a Transformer-enhanced Atrous Spatial Pyramid Pooling module to capture long-range contextual information, while MobileNetV2 is employed as the backbone to reduce model parameters and improve computational efficiency. In addition, the Convolutional Block Attention Module is introduced to enhance spatial and channel feature representation. A manually annotated rice field dataset was constructed for model training and evaluation. Experimental results demonstrate that RFE-Net provides an effective balance between segmentation accuracy and computational efficiency, achieving an mIoU of 80.52%, Precision of 89.46%, Recall of 89.03%, F1-score of 89.19%, and OA of 89.33%. These results indicate that RFE-Net is an effective approach for accurate and efficient rice field extraction from remote sensing imagery.