TerraLeaf-Net: Fusing Multispectral Satellite Imagery and Vision Transformers for Early Detection of Wheat Rust Disease at Field Scale

Wheat rust, caused by fungi of the genus Puccinia, is among the most economically devastating crop diseases globally, capable of reducing yields by up to 70% in severe epidemics. Early detection at field scale is essential for timely fungicide intervention, yet conventional ground scouting is labour-intensive and cannot scale to the regional monitoring required for epidemic management. This paper presents TerraLeaf-Net, a deep learning framework that fuses multispectral Sentinel-2 satellite imagery with a vision transformer backbone to detect and map wheat rust infection at 10-metre resolution before symptoms are visible to the naked eye. We construct RustSat-4K, a new georeferenced dataset of 4,200 field patches spanning three growing seasons across wheat-producing regions in Morocco and Colombia, each labelled with ground-truth disease severity collected by agronomists. TerraLeaf-Net combines a spectral-index feature stream (incorporating NDVI, the Red-Edge Chlorophyll Index, and a custom Rust Stress Index) with a Swin Transformer spatial encoder through a cross-attention fusion module. On the held-out test set, TerraLeaf-Net achieves a four-class severity classification accuracy of 91.4% and macro-F1 of 0.887, outperforming a ResNet-50 CNN baseline by 6.2% F1 and a spectral-index random forest by 13.9% F1. Critically, the model detects pre-symptomatic infection (asymptomatic but infected fields) with 0.81 recall, offering a 7-10 day early-warning window over visual scouting. Ablation confirms that multispectral-spatial fusion contributes 4.1% F1 over imagery alone.