Discriminating rice phenological stages using UAV multispectral vegetation index statistics at field-block scale

Accurate discrimination of rice phenological stages is important for irrigation scheduling, fertilizer timing, and field-level crop management. This study evaluated whether field-block statistics derived from unmanned aerial vehicle (UAV) multispectral imagery could discriminate observed rice stages in fragmented tropical irrigated fields. Imagery was acquired using a fixed-wing vertical-take-off-and-landing UAV with a five-band sensor. Spectral information within field-block polygons was represented by vegetation-index statistics and reflectance-scaled five-band descriptors. Five predefined feature configurations were compared using Random Forest, support vector machine, and linear discriminant analysis. Median imputation, low-variance removal, Pearson-correlation filtering at |r| >0.92, and standardization were fitted independently within each training fold. Model evaluation used stratified group cross-validation by field block. The dataset comprised 93 field-block observations (19 planting, 46 tillering, and 28 booting). The 36-variable four-core-vegetation-index configuration provided the strongest field-block generalization, outperforming five-band descriptors alone. Random Forest achieved an accuracy of 0.8817, Cohen’s kappa of 0.8059, macro-F1 of 0.8708, and weighted-F1 of 0.8807. Permutation importance identified enhanced vegetation index mean (EVI_mean), two-band enhanced vegetation index median (EVI2_median), and EVI2_mean as the dominant predictors. Compact vegetation-index statistics provided an effective field-block representation within the observed landscape, although the analysis covered only three stages of the BBCH scale.