The research, led by Shivranjani Baruah of Cornell University together with collaborators at Cornell, the International Crops Research Institute for the Semi-Arid Tropics (ICRISAT) in Hyderabad, and the Indian Institute of Rice Research, set out to answer a deceptively simple question: can a four-band multispectral sensor paired with 3D laser scanning classify bacterial blight severity as reliably as the expensive hyperspectral systems that dominate the field? The answer, in most cases, is yes — provided the data are processed in the right way.

The team conducted their experiments on the LeasyScan platform at ICRISAT, an outdoor high-throughput phenotyping facility that moves plants past a Phenospex sensor combining LiDAR-based 3D reconstruction with RGB and near-infrared imaging. Ten rice genotypes were crossed with four Xoo bacterial strains, producing 17 unique host-pathogen combinations that spanned the full spectrum of disease outcomes, from complete resistance to extreme susceptibility. The panel included near-isogenic IRBB lines each carrying a single known resistance gene — executor genes, NLR proteins, receptor-like kinases, and recessive alleles of susceptibility genes — alongside four popular southern Indian cultivars that serve as susceptible checks. After clip inoculation, plants were scanned daily from the day of inoculation until 16 days post inoculation, while two independent plant pathologists scored disease severity on the standard six-point scale derived from the International Rice Research Institute’s Standard Evaluation System.

The resulting dataset of 3,073 3D point clouds became the raw material for a series of machine learning experiments. The key insight came from how the spectral data were handled. Conventional approaches typically average vegetation indices — mathematical combinations of reflectance values that track chlorophyll content, pigment degradation, and senescence — across an entire plant canopy. But bacterial blight does not attack a canopy uniformly. It begins as water-soaked lesions that expand into necrotic patches, leaving healthy and diseased tissue intermingled on the same leaves. Averaging, the researchers reasoned, throws away exactly the information that matters most.

To test this, the team developed two complementary ways of preserving within-canopy spectral heterogeneity. The first was supervised binning: for each vegetation index, five thresholds were optimized using simulated annealing to divide continuous index values into six bins, with the objective of maximizing the separation between disease severity classes. Each point cloud was then summarized by the proportion of points falling into each bin. The second was unsupervised: k-means++ clustering with 12 clusters partitioned point-level spectral features into recurrent spectral states, and each plant was described by the relative abundance of each state within its canopy. Kernel density estimation confirmed that vegetation indices such as NDVI, SRI, and PSRI shifted systematically with severity — NDVI, for example, declined from 0.558 in healthy plants to 0.500 in highly susceptible ones — and Kolmogorov-Smirnov tests showed significant differences among all severity classes for every index examined.

The results were striking. When severity was classified into four broad categories — healthy, moderately resistant, susceptible, and highly susceptible — a random forest model trained on binned vegetation indices alone achieved 76.2 percent accuracy on an independent holdout test set, with a weighted F1 score of 0.76 and a quadratic weighted kappa of 0.86, a measure that penalizes large misclassifications more heavily than small ones. The mean absolute error was just 0.28 class units. Adding cluster-ratio features or structural traits from the 3D scans did not meaningfully improve on this. By contrast, models built on conventional canopy-average vegetation indices lagged behind, reaching only 69.7 percent accuracy. The gap widened when the task was made harder: for the original six-category scale, the best average-index model managed just 58 percent accuracy, while a support vector machine combining cluster ratios and binned indices reached 71 percent.

The pattern held across algorithms. XGBoost performed poorly in every configuration, while random forest and support vector machines traded the lead depending on the classification task. Feature importance analysis revealed that binned indices such as NGBI, HUE, PSRI, GRVI, and VARI dominated the top predictors, alongside a handful of cluster-ratio variables. Notably, the models were never told which genotype, bacterial strain, or day post inoculation each scan came from — they had to infer severity purely from the spectral and structural signatures of the plants themselves. Error analysis showed that misclassifications clustered in specific genotypes, particularly the resistant lines IR10, IR21, and IR5, and were concentrated between neighboring severity classes, suggesting that adjacent categories form a biological continuum rather than sharply separable states.

That finding has practical implications for breeding programs. If fine distinctions between neighboring severity scores are inherently unstable — influenced by lesion distribution, symptom timing, lighting, canopy architecture, and scorer subjectivity — then selection decisions may benefit more from robustly separating broad resistance classes than from chasing unreliable precision on a fine scale. The four-class framework delivered exactly that: reliable discrimination between healthy, intermediate, and susceptible responses, with the healthiest class classified with 93 percent recall and the most susceptible at 78 percent.

What makes the study genuinely exciting is its economics. Hyperspectral imaging systems, which measure reflectance across dozens or hundreds of narrow wavelength bands, can cost prohibitive amounts and demand serious computational resources — barriers that put them out of reach for many breeding programs in the very regions where rice disease pressure is highest. This work demonstrates that a simple four-band sensor, the kind already mounted on consumer drones and agricultural scanners, can deliver usable classification accuracy if paired with 3D structural context and heterogeneity-preserving data processing. The 3D component matters because it lets researchers interpret spectral signals in their spatial context within the canopy, capturing subtle changes in leaf angle and leaf area that flat imaging misses.

The authors are careful to note the limitations. Severity labels themselves carry uncertainty, especially for intermediate classes where visual assessment is subjective. The models do not yet exploit explicit lesion localization or temporal progression, and some degree of confounding between disease severity and plant developmental stage is possible. Broader validation across seasons, environments, and genetic backgrounds will be needed before the framework becomes a routine breeding tool. The team has made its models, processed datasets, and analysis pipelines publicly available to accelerate that process. But the direction of travel is clear: as laser and depth sensors become cheaper and more portable, the combination of low-dimensional spectral sensing with 3D phenomics could bring objective, high-throughput disease screening within reach of the breeding programs that need it most — and not just for rice. The framework could be adapted for foliar diseases across many crops, potentially accelerating resistance breeding at a moment when food security under climate stress has never mattered more.

Subject of Research: 3D multispectral phenotyping for rice bacterial blight severity classification

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

Article References: Baruah, S., Du, R., Kumar, D., Laha, G., Choudhary, S., Kholova, J., Jiang, Y., Bogdanove, A. J., & Gold, K. M. (2026). A practical framework for rice bacterial blight severity classification using 3D multispectral sensing. Smart Agricultural Technology, 15, Article 102586. https://doi.org/10.1016/j.atech.2026.102586

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102586

Keywords: rice, bacterial blight, Xanthomonas oryzae, 3D phenotyping, multispectral imaging, machine learning, vegetation indices, plant breeding, disease severity, LiDAR, food security, precision agriculture