Agricultural researchers tracking fungal epidemics in staple grains have demonstrated that unmanned aerial vehicle multispectral remote sensing combined with targeted machine learning feature selection can accurately identify and map rice blast disease across varied growing stages and fertilization levels. Conducting repeated field trials over consecutive growing seasons in 2019 and 2020 in Wanchang Town, Yongji County, Jilin Province, China, the research team examined canopy-level spectral changes in japonica rice exposed to the fungal pathogen Magnaporthe oryzae. Rice blast presents an ongoing hazard to global cereal supplies, causing typical annual yield losses of 10 to 30 percent during major outbreaks and threatening complete crop destruction in severe field epidemics. While traditional diagnostic workflows like manual field walking, enzyme-linked immunosorbent assays, and polymerase chain reaction assays remain reliable reference benchmarks, their operational costs, labor demands, and sampling delays restrict rapid deployment over extensive agricultural areas. By evaluating drone imagery paired with tailored analytical pipelines, the investigators sought to establish an objective, scalable framework for quantifying canopy symptoms and guiding timely intervention.

The study focused on the representative cold-temperate japonica rice cultivar Jijing 88, which exhibits a 150-day total growth period and moderate to high field resistance to blast infection. To initiate realistic epidemic dynamics without introducing artificial plot-level inoculum gradients directly onto the target cultivar, the researchers planted the highly blast-susceptible cultivar Mongolian rice as an infector spreader along designated experimental rows. A conidial suspension of Pyricularia oryzae was applied directly to the Mongolian rice leaves at the tillering stage using a hand sprayer, after which spores dispersed naturally to the surrounding Jijing 88 canopy through wind and rain splash. Because elevated nitrogen fertilization promotes dense canopy architecture and alters tissue chlorophyll concentrations, the experimental layout allowed the researchers to assess how varying nutritional baselines influence canopy reflectance alongside fungal pathology across the booting, heading, and milk development stages.

Ground-truth disease severity was cataloged systematically during each aerial campaign by inspecting sampled hills according to national technical standards for blast forecasting. Agronomists determined the disease incidence rate by dividing the count of symptomatic leaves by total leaves per surveyed hill, sorting plots into six standardized severity grades spanning from grade zero for completely healthy foliage to grade five for leaf surface damage exceeding 50 percent necrosis. Across both test years, ground evaluations recorded disease progression advancing from booting into heading and milk stages, with the total number of symptomatic sample units expanding markedly as the epidemic developed. In parallel, aerial data collection missions were conducted at 11:00 a.m. local time under clear atmospheric conditions using a DJI M600 Pro multi-rotor aerial platform carrying a RedEdge-MX multispectral camera at a stabilized flight altitude of 50 meters above ground level. The sensor acquired narrow-band imagery across blue (475 nm), green (560 nm), red (668 nm), red edge (717 nm), and near-infrared (840 nm) wavebands with forward overlap of 60 percent and side overlap of 50 percent to support seamless orthomosaic generation.

To convert raw sensor digital numbers into dependable physical measurements unaffected by solar illumination drift, the researchers collected calibration panel imagery immediately before and after each flight mission. Orthorectified reflectance mosaics generated through Pix4Dmapper were normalized band-by-band to minimize ambient lighting variability before computing downstream canopy descriptors. Pathogen infection fundamentally disrupts host physiological structures, degrading green leaf pigments and collapsing mesophyll cell integrity. At the booting stage, when visual blast symptoms remained mild across all experimental plots, canopy reflectance differences were governed predominantly by nitrogen rates, showing progressive increases in green and red edge bands alongside sustained high near-infrared scattering. However, as disease severity intensified during heading and milk stages, fungal damage superseded nutritional effects, producing an observable reduction in near-infrared reflectance alongside elevated or flattened reflectance across the red and red edge spectral domains.

Recognizing that the contrasting behavior between red edge absorption and near-infrared scattering serves as a diagnostic indicator of infection, the authors designed three modified red edge vegetation indices: RRNDVI, RRDVI, and RRRVI. These custom formulations substituted the standard near-infrared band in conventional normalized difference, difference, and ratio vegetation indices with the narrower red edge waveband centered at 717 nanometers. When evaluating correlations between individual features and observed severity ratings, the red band exhibited the strongest independent linear association, followed closely by the red edge band, while the near-infrared band displayed the weakest direct correlation. The newly introduced red edge indices demonstrated strong negative associations with disease severity (|r| ≥ 0.2), confirming that mathematically isolating red edge responsiveness helps suppress background soil interference and illumination anomalies while directly tracking chlorophyll degradation. Complementing the spectral bands and ten established vegetation indices, the authors calculated eight gray-level co-occurrence matrix texture metrics across each image band in ENVI software to capture spatial structural variations.

The evaluation of single-feature inputs revealed clear performance differences across four machine learning algorithms: Random Forest, Support Vector Regression, Partial Least Squares Regression, and Gradient Boosting Decision Trees. Hyperparameters for tree and kernel algorithms were refined using grid search and five-fold cross-validation on an 80:20 training-testing split, while Partial Least Squares Regression was evaluated without hyperparameter optimization to prevent overfitting. When using single feature categories alone, vegetation indices generally delivered the highest diagnostic accuracy, followed by raw spectral bands and texture metrics. Among all single-input models, Gradient Boosting Decision Trees utilizing raw spectral bands yielded a coefficient of determination of 0.732 and a root mean square error of 0.998, whereas Partial Least Squares Regression relying on texture metrics performed least effectively, producing an R² of 0.283 and a root mean square error of 1.633. Random Forest maintained the most consistent single-input stability across all feature sets, with coefficients of determination consistently exceeding 0.550.

Integrating multi-source input features into combined modeling pipelines further elevated diagnostic performance across most architectures. The research team evaluated four distinct data combinations: spectral reflectance with vegetation indices (C1), spectral reflectance with texture (C2), vegetation indices with texture (C3), and the full pool of spectral, index, and texture metrics (C4). Random Forest incorporating vegetation indices and texture metrics (C3) attained an R² of 0.795, reflecting an 8.6 percent improvement in explanatory power and a 12.6 percent reduction in root mean square error to 0.873 compared to the best single-feature Random Forest model. Similarly, Gradient Boosting Decision Trees reached an R² of 0.787 and root mean square error of 0.904 when pairing spectral reflectance with texture features (C2). However, pairing all available features simultaneously under combination C4 did not consistently outperform more compact subsets, indicating that uncontrolled feature expansion introduces data redundancy and noise that can impede model convergence.

To mitigate the risks of feature redundancy and overfitting, the authors systematically benchmarked filter, wrapper, and embedded feature selection techniques. Filter selection based on Pearson correlation ranking and texture contribution analysis provided substantial performance gains across linear and kernel-based algorithms; for example, applying filter selection to the C3 feature set increased the R² of Partial Least Squares Regression from 0.498 to 0.751, an improvement of 50.8 percent. Embedded regularization methods, including LASSO and Elastic Net, proved advantageous for Support Vector Regression and Partial Least Squares Regression by penalizing collinear predictors, though they frequently suppressed performance when applied to decision tree ensembles. Conversely, Recursive Feature Elimination acting as a wrapper method aligned exceptionally well with tree-based models, allowing Random Forest and Gradient Boosting Decision Trees to maintain steady diagnostic precision even after eliminating up to 79 percent of candidate variables. With wrapper-optimized inputs, Gradient Boosting Decision Trees achieved the highest overall diagnostic accuracy in the study, reaching an R² of 0.850 and a root mean square error of 0.748 using 22 selected features from combination C1.

Interpretability assessments using SHAP (Shapley Additive Explanations) provided biophysical context for these algorithmic outcomes. Circular network and interaction matrix analyses demonstrated that the Structure Insensitive Pigment Index (SIPI) registered the highest overall feature importance, followed by indices such as NGRDI, VARI, RVI, NDVI, NDRE, and direct red edge reflectance. Dense interaction networks formed between pigment-sensitive indices, the near-infrared band, and visible wavebands, confirming that blast damage is best captured through non-linear couplings that reflect concurrent pigment loss and foliage structural breakdown. While most texture metrics exhibited negligible standalone importance, red edge texture features displayed noticeable synergistic interactions, providing spatial context that helped models separate patchy lesions from canopy background noise. Tracking feature weights before and after selection confirmed that the filtering workflows successfully pruned uninformative variables while concentrating predictive weighting on key disease-sensitive spectral indices.

Cross-period validation revealed important operational strengths and practical boundaries when deploying models trained on full-season data (booting, heading, and milk stages) to earlier growth phases. When testing models exclusively on booting-stage data, overall predictive accuracy declined relative to full-season testing, reflecting the analytical challenge of isolating subtle early lesions from canopy variations driven by baseline nitrogen fertilization. Nevertheless, tree-based models maintained stable predictive sensitivity, indicating that core red edge traits retain diagnostic utility for timely field alerts. Aerial severity mapping on August 21 and August 30, 2020, confirmed these dynamics across the field landscape: initial infections clustered around the spreader rows and lower field margins before spreading rapidly toward central high-nitrogen plots and downwind northeastern rows due to prevailing summer winds. Although the field trial was restricted to a single cultivar without isolated single-factor inoculation controls, the resulting integration of drone multispectral sensing and tailored feature selection establishes a robust quantitative methodology for precision disease surveillance in grain production.

Subject of Research: UAV-based multispectral remote sensing and machine learning feature selection for rice blast disease diagnosis

Article Title: Diagnosis of rice blast disease using UAV-based multispectral data: a comparative study of feature selection methods

Article References: Wu, Y., Song, S., Zhu, B., & Song, K. (2026). Diagnosis of rice blast disease using UAV-based multispectral data: a comparative study of feature selection methods. Smart Agricultural Technology, 15, Article 102701. https://doi.org/10.1016/j.atech.2026.102701

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102701

Keywords: Rice Blast, Magnaporthe oryzae, UAV Remote Sensing, Multispectral Imaging, Feature Selection, Precision Agriculture, Red Edge Vegetation Indices, Random Forest, Gradient Boosting Decision Tree, Recursive Feature Elimination, Japonica Rice, Crop Disease Diagnosis, Canopy Reflectance, Nitrogen Gradient, Plant Pathology, Smart Farming, Agricultural Remote Sensing