Synergistic Monitoring of Cotton Verticillium Wilt Using GF-2 Satellite and UAV Data: A Dual-Branch Feature Fusion Approach

• A dual-track fusion framework integrating pixel-level and decision-level strategies was developed for field-scale monitoring of cotton Verticillium wilt using GF-2 satellite and UAV multispectral imagery. • The PS-based pixel-level fusion method outperformed GS and PC approaches in preserving spectral fidelity (SAM = 0.362) and spatial structure (SSIM = 0.967), providing a robust data foundation for disease characterization. • Coupled physiological-structural features (VITF), combining vegetation indices and texture metrics, significantly enhanced the separability of disease severity gradients under complex field conditions. • Metaheuristic optimization algorithms (PSO, ACO, GWO) substantially improved model performance, with PSO-ELM achieving the highest inversion accuracy (R 2 =0.829) within the pixel-level fusion framework. • Decision-level fusion via performance-weighted averaging of UAV and satellite models yielded the best overall accuracy (R 2 =0.8372, RMSE = 3.63), demonstrating the value of cross-scale radiometric complementarity over simple spatial enhancement. • Systematic comparison of fusion paradigms revealed that pixel-level fusion is for fine-scale mapping, while decision-level fusion offers greater robustness and scalability for regional disease monitoring and operational applications. This study focused on field-scale remote sensing–based monitoring of cotton Verticillium wilt, addressing two key challenges: limited inter-class spectral separability and strong intra-class heterogeneity, as well as the synergistic mechanisms through which multisource data integration can mitigate these limitations. To this end, a dual-track analytical framework that integrates pixel-level and decision-level fusion was developed. By jointly exploiting wide-swath Gaofen-2 satellite imagery and high-resolution multispectral data acquired by a DJI M600 Pro unmanned aerial vehicle (UAV), this study systematically elucidated the end-to-end mechanism spanning data synergy, feature construction, model optimization, and spatial extrapolation. The results indicate that pan-sharpening (PS) constitutes the most effective data foundation for subsequent modeling along the pixel-level fusion path. The PS method consistently outperforms the Gram–Schmidt and principal component analysis approaches in both spectral fidelity (SAP = 0.362) and spatial detail preservation (SSIM = 0.967). At the feature representation level, the coupled “physiological–structural” feature set, which integrates vegetation indices and texture features, exhibited superior robustness in characterizing disease severity gradients. Within the particle swarm optimization (PSO)–extreme learning machine (ELM) model, the coefficient of determination (R²) for severely affected areas improved from 0.763 to 0.829. Within the pixel-level fusion framework, the GWO–SVM model exhibited robust performance (R² = 0.7443), whereas under the decision-level fusion framework, the PSO–ELM model achieved a higher R² = 0.8104. Furthermore, a weighted-average decision-level fusion strategy, namely UAV–satellite wide-field synergistic fusion, produced the highest inversion accuracy (R² = 0.8372, mean absolute error = 2.90, and root mean square error = 3.63). This improvement cannot be attributed solely to spatial resolution enhancement; rather, it arises from the “soft constraint” imposed by the radiometric stability of satellite observations on UAV-derived spectral noise, in addition to the complementarity of cross-scale information. Clear performance demarcations between the two fusion paths were identified: pixel-level fusion was more appropriate for fine-scale mapping in targeted areas, whereas decision-level fusion, owing to its modular structure and cross-platform compatibility, was better suited for large-area risk screening, operational monitoring, and uncertainty characterization. Overall, this study offers both methodological support and theoretical insights for achieving high-accuracy and scalable remote sensing–based inversion of cotton Verticillium wilt.