Multi-temporal UAV multispectral phenotyping and machine learning for modeling cotton defoliation dynamics across diverse accessions

Cotton defoliation is critical for efficient mechanical harvesting, yet its temporal progression across diverse germplasm remains difficult to quantify at high throughput. This study developed a multi-temporal unmanned aerial vehicle (UAV)-based machine-learning framework to characterize post-treatment defoliation dynamics across 419 cotton accessions. Multispectral imagery acquired before and after defoliant application was used to derive spectral and texture features, and an Entropy Weight Method (EWM)-based strategy was used for feature ranking and selection. Multiple regression algorithms were evaluated using accession-grouped cross-validation. Spectral indices provided substantially stronger predictive information than texture features, while relative changes from the pre-treatment baseline further improved model performance. Temporal encoding provided a modest but complementary contribution. Among the evaluated algorithms, Random Forest achieved the best overall performance, with an R² of 0.73 and an RMSE of 10.69%. Accession-level trajectory analysis subsequently identified four post-treatment defoliation patterns, with variation primarily reflected in the magnitude and rate of defoliation progression. These results demonstrate the potential of multi-temporal UAV phenotyping for scalable characterization of cotton defoliation dynamics and provide a practical basis for germplasm evaluation and mechanized harvest management.