A partially supervised joint segmentation method for farmland and rural roads from single-class annotated remote sensing datasets

Abstract

Accurate segmentation of co-existing land-cover objects in high-resolution remote sensing imagery is essential for precision agriculture and rural planning. Although multi-class datasets are increasingly available, many high-quality annotations remain task-specific, where only a single target class is labeled and other co-occurring objects are treated as background. Jointly exploiting such heterogeneous single-class datasets introduces supervision ambiguity, resulting in pseudo-background interference, feature conflicts, and unstable optimization. To address these challenges, we propose RSFRNet, a partially supervised joint segmentation framework for farmland and rural road extraction from mutually exclusive single-class annotations. The key idea of RSFRNet is to explicitly match different types of conflicts in this setting with corresponding network and optimization designs. First, considering that roads are dominated by thin linear structures whereas farmland is characterized by large homogeneous regions, we design a Filtering-based Feature-wise Linear Modulation module (F-FiLM) to adapt decoder features to class-specific structural patterns. Second, to reduce the interference caused by unlabeled co-existing objects being treated as background, we introduce a Learnable Class Codebook (L-Codebook), which uses class prototypes to refine encoder representations before decoding. Third, to alleviate inconsistent parameter updates caused by heterogeneous single-class supervision, we introduce PSL-PCGrad to correct conflicting gradients during joint training. In addition, task-aware loss constraints are used to preserve road connectivity and farmland regional integrity. In this way, RSFRNet addresses partially supervised farmland–road joint segmentation from feature adaptation, semantic purification, and optimization stabilization perspectives. Experiments on FarmSeg-VL and WHU-RuR + show that RSFRNet achieves the best evaluated performance, reaching 74.04 % mIoU with balanced farmland and road IoU. These results highlight the effectiveness of RSFRNet for partially supervised farmland–road joint segmentation under mutually exclusive single-class annotations and suggest its potential for related single-class annotation scenarios.