Multi-source UAV remote sensing for cotton Verticillium wilt resistance grading using TPAC-Net.

Introduction: Cotton Verticillium wilt resistance grading under field conditions remains challenging because disease symptoms are spatially heterogeneous within plots, local pathological cues are unevenly distributed across high-resolution unmanned aerial vehicle (UAV) orthomosaics, and disease-related responses are not expressed with the same sensitivity or spatial consistency across RGB, raw multispectral (MS), and vegetation-index (VI) representations. Methods: To address these challenges, this study proposes TPAC-Net, a UAV-based Tri-source Pathological Alignment and Coupling Network for plot-level cotton Verticillium wilt resistance grading, which organizes red-green-blue (RGB) imagery, raw MS imagery, and VI representations derived from MS bands as complementary representation streams of visible canopy structure, band-level spectral responses, and index-enhanced disease-sensitive spectral contrasts. Based on field survey records and the relative disease index (RDI) defined by GB/T 22101.5-2009, two-class and five-class classification settings were constructed, with the standardized five-class task serving as the primary benchmark. TPAC-Net combines source-specific encoders, a Tri-Source Feature Attention Module (TFAM) for adaptive pathological feature coupling, a Cross-Source Semantic Alignment Module (CSAM) for high-level semantic consistency among complementary representations, and a Local Pathogenic Patch Perception strategy for plot-level decision-making from local patches. Results: The proposed framework achieved F1-scores of 79.20% and 50.37% under the two-class and five-class settings, respectively. Under the primary five-class task, it yielded an absolute F1-score gain of 14.36 percentage points over the best-performing baseline. Localized-response visualization showed that higher model responses were concentrated in selected local canopy regions, although fine-grained discrimination among intermediate resistance grades remained challenging. Discussion: These results suggest that multi-source UAV remote sensing can support field-survey-derived RDI resistance-grade inference through structured representation alignment, pathological feature coupling, and local evidence aggregation.