An improved SegFormer for remote sensing segmentation of construction and demolition waste via frequency-spatial joint perception

Accurately segmenting construction and demolition waste (CDW) in high-resolution remote-sensing images is challenging because of spectral confusion, scattered distributions, and irregular boundaries. We propose a frequency-spatial joint-perception framework based on SegFormer. The original decoder is replaced with an Enhanced Multi-Scale Feature Fusion Network (E-MSFM), which contains a Spectral-Spatial Context Module (SSCM) for multi-scale feature interaction and a Spectral-Aware Boundary Enhancement (SABE) module for preserving high-frequency detail and delineating irregular boundaries. A weighted Focal-Dice loss is used to reduce missed detections of small targets under severe class imbalance. On the Construction Waste Landfill Dataset (CWLD), the proposed method achieved an mIoU of 90.57%, exceeding the original SegFormer and DeepLabV3+ by 3.20 and 3.45% points, respectively. These results demonstrate that frequency-spatial joint perception can effectively improve CDW segmentation in high-resolution remote-sensing imagery.