Fusion of Multisensor and Multitemporal Data for Early Diagnosis of Litchi Downy Blight Disease

Aiming at the problem of multisensor fusion, this article takes litchi downy blight (LDB) disease as the research object. The disease is influenced by multiple factors, making IoT environmental data collection and analysis vital for disease prediction. Hyperspectral imaging, a noninvasive sensing technology, can detect changes in litchi leaf disease progression. Multispectral remote sensing offers exceptional real-time and large-scale monitoring capabilities for crop disease surveillance and early warning, quickly acquiring comprehensive orchard health data. Environmental data, hyperspectral data, and multitemporal multispectral imagery were acquired through IoT sensors, ground-based spectrometers, and uncrewed aerial vehicle (UAV)-mounted multispectral cameras, respectively. In order to overcome the limitations of single-source data for optimal detection timing and enhance early disease monitoring accuracy and reliability, we introduced a cross-attention mechanism-based multimodal feature fusion network (CAMF-Net). During feature extraction, the gray-level co-occurrence matrix (GLCM) extracted texture features from multispectral images, the successive projections algorithm (SPA) extracted hyperspectral feature bands, and the Monte Carlo algorithm combined with the Pearson correlation coefficient analyzed sensitive environmental factors. By leveraging the fused texture features, hyperspectral bands, and environmental factors, we conducted comparative experiments with seven mainstream models to demonstrate the effectiveness and superiority of the proposed method. The results indicated that, in the testing experiments conducted on the self-built dataset, the average F1 Score and average. Accuracy of CAMF-Net reaches 91.88% and 92.68%, respectively.