SPAF-SegNeXt: Prior-Guided Attention and Frequency-Spatial Hybrid Fusion for Multimodal Sugarcane Segmentation in Remote Sensing
1939-1404
Abstract
Accurate monitoring of the spatial extent of sugarcane cultivation is critical for optimizing land-use efficiency and regulating the sugar industry. However, sugarcane cultivation areas are characterized by complex topography and fragmented plots, compounded by frequent occurrences of “same object with different spectra” and “different objects with similar spectra,” posing significant challenges for remote sensing-based extraction. Existing methods are limited by insufficient spectral-topographic coupling, loss of high-frequency boundary details, and boundary drift. To address these challenges, we propose a novel multimodal hybrid segmentation framework, termed Sugarcane Prior-guided Attention and Frequency-fused SegNeXt (SPAF-SegNeXt), for end-to-end semantic segmentation of sugarcane fields. First, we construct a specialized multimodal dataset, the Sugarcane Augmented Remote-sensing Fusion with Topographic Prior Dataset (SugarARF-Topo), covering diverse geomorphological scenarios, with high-quality annotations ensured by a dual-track strategy combining visual interpretation and data-driven validation. Second, we design the Sugarcane Prior-guided Attention Module (SugarPriorAM), integrating agronomic threshold-based prior knowledge with deep learning to provide spatial guidance, suppress background noise, and enhance target feature representation. In addition, we incorporate an existing Frequency-Aware Feature Fusion (FreqFusion) module to integrate frequency-domain information, alleviating intraclass inconsistency and mitigating information loss due to downsampling. Finally, by integrating these two modules, we construct the SPAF-SegNeXt architecture, following a “threshold-prior guidance–frequency-domain feature fusion–refined segmentation” pipeline, and conduct end-to-end experiments on the SugarARF-Topo dataset. Results demonstrate that SPAF-SegNeXt achieves superior performance, with an IoU of 90.13%, significantly outperforming existing methods. This framework provides an effective solution for accurate monitoring of sugarcane cultivation spatial extent, offering both theoretical and practical support for land-use optimization and sustainable sugar industry management.