Causal Stress Deconvolution for Precision Soybean Management: A Multi-Modal Deep Learning Approach Integrating Topography, Climate, and Biophysical Indices

The stability of global soybean production is challenged by the diagnostic limitations of conventional remote sensing, which relies on symptomatic spectral indices (e.g., normalized difference vegetation index, NDVI) that cannot distinguish between spectrally similar but etiologically distinct stressors.To resolve this ambiguity, this study proposes the causal stress deconvolution network (CSDN), a novel deep learning framework that transitions diagnosis from reactive symptom monitoring to physicallyconstrained causal inference.Unlike standard correlational models, the CSDN explicitly structures the diagnostic process by conditioning symptomatic biophysical data (green leaf area index, gLAI; normalized difference red edge, NDRE) on two mechanistic environmental drivers: the topographic wetness index (TWI) as a static spatial filter, and accumulated precipitation (AP) as a dynamic temporal trigger.Validated on a soybean cultivation site in South Korea with complex topography, the CSDN achieved a diagnostic F1-score of 0.978 for wet-soil disease (e.g., Phytophthora rot), significantly outperforming a symptomonly baseline (F1 = 0.959).Crucially, a rigorous ablation study confirmed that the integration of the static TWI filter provided the single largest performance gain (Absolute F1 = +0.015),effectively eliminating false positives in low-risk zones.By mechanically linking spectral decline to its hydrological preconditions, this research demonstrates that integrating physical geography with deep learning is essential for resolving diagnostic uncertainty in precision agriculture.