Work

A practical framework for rice bacterial blight severity classification using 3D multispectral sensing

Year: 2026

Type: article

Abstract: Assessing crop disease severity is difficult under field conditions because disease symptoms develop progressively, and vary spatially, within and among plants. Precision breeding programs therefore require phenotyping approaches that can resolve variation in symptoms while remaining scalable under field conditions. While imaging spectroscopy has advanced high-throughput stress detection, translating these tools to field environments remains difficult because high sensor cost, slow acquisition over large plots, variable illumination and background conditions, complex data-processing requirements, and limited model generalization across genotypes and environments constrain routine deployment. This paper presents a multimodal, high-throughput phenotyping framework to classify bacterial leaf blight severity in rice, by integrating multispectral imaging with 3D structural sensing in a semi-controlled field setting. Across 10 rice genotypes with varying susceptibility to Xanthomonas oryzae pv. oryzae, disease-driven variation was captured using spectral vegetation indices, color-based metrics, and structural traits derived from point cloud data. To better capture within-canopy heterogeneity, vegetation indices derived from point cloud data were summarized using both supervised spatial binning and unsupervised k-means++ clustering to generate features representing distinct canopy states. These features were integrated with 3D structural traits for disease severity classification. Among the evaluated models, random forest models trained on binned vegetation indices, either alone or combined with spectral cluster features, achieved the strongest performance for four-category disease severity classification, with an accuracy of 0.762, compared with 0.697 for a model trained only on canopy-average vegetation indices. This approach improves the interpretability of disease signals and reduces sensitivity to within-plot variability. These results demonstrate that preserving within-canopy spectral heterogeneity using low-dimensional RGB-NIR and 3D point-cloud-derived features improves disease severity quantification, and offers a practical alternative to high-dimensional hyperspectral approaches.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2026.102586