Early detection of Brown Spot disease ( Bipolaris oryzae ) in tropical rice systems is constrained by both operational and spectral limitations that hinder large-scale implementation. In-field, the standard Fedearroz protocol, based on W-pattern systematic sampling conducted by an expert agronomist, becomes impractical over large production areas such as those in Casanare. At the remote sensing scale, UAV flights at conventional altitudes ( ≥ 10 m) also fail to resolve the problem, as phenotypic ambiguity of early lesions and spectral mixing with the canopy background reduce detectability. Under these conditions, conventional monitoring approaches fail due to the saturation of structural indices in dense vegetation. We propose a framework that integrates proximal multispectral sensing ( h = 40 –50 cm) with a deep learning architecture, based on the premise that spatial resolution governs pathogen detectability. Through a high-throughput audit of 17 multispectral indices, we constructed a high-dimensional Bio-Spectral Tensor combining CVI, MACI, and GNDVI, selected for their complementary capacity to capture distinct physiological signatures. Fisher Discriminant analysis showed that CVI increases class separability by a factor of 144 relative to NDVI, effectively decoupling biotic stress signals from environmental artifacts. These tensors were processed using the ConvNeXt architecture, achieving an overall accuracy of 98.95% and a Recall of 1.00 for the diseased class, with zero false negatives in independent evaluation. Grad-CAM analysis revealed that model attention is concentrated on necrotic centers and chlorotic halos, confirming the pathophysiological grounding of the predictions. Taken together, these results establish a physically grounded framework for the early detection of Bipolaris oryzae in tropical rice systems, identifying h = 40 –50 cm as the critical spatial threshold for supervised lesion discrimination and defining a scalability roadmap toward operational UAV deployment via high-resolution multispectral sensors.
M et al. (Mon,) studied this question.