UAV-based multispectral imaging and machine learning for characterizing herbicide injury and desiccant responses in field crops

Abstract Purpose Herbicides and desiccants are critical for weed control and harvest management; however, their efficacy varies across crop species, environments, and growing seasons. This study evaluated unmanned aerial vehicle (UAV)-based multispectral imaging combined with machine-learning (ML) as a non-destructive approach for quantifying herbicide injury and desiccation responses across Canadian Prairie agroecosystems. Methods Field trials were conducted from 2022 to 2024 at three sites in Alberta and Saskatchewanusing tame oat, oriental mustard, and lentil. Tame oat and oriental mustard were used as pseudo-weed models for wild oat and wild mustard, respectively. UAV multispectral imagery and concurrent ground-based visual ratings were collected from 0 to 21 days aftertreatment to characterize temporal crop responses. Six ML models were evaluated using UAV-derived vegetation indices. Results Ensemble models, including Gradient Boosting, Extreme Gradient Boosting, and Random Forest, consistently demonstrated strong predictive performance, with site-specific coefficients of determination (R) of 66–90% in oat, 60–75% in mustard, and 43–96% in lentil. Multi-site models achieved R values of 65–83%; however, leave-one-site-out (LOSO) and leave-one-year-out (LOYO) validation for oat and mustard showed reduced predictive performance in independent environments, highlighting limitations in model transferability. Vegetation indices captured temporal herbicide injury and stress dynamics across crops, while area under the desiccation progress curve (AUDPC) analysis quantified lentil dry-down and differentiated desiccant performance. Pearson correlation analysis identified NDVI and NDGI indices strongly associated with visual injury ratings across crops and environments. NDVI showed the most consistent relationships in oat and mustard, whereas NDVI, NDGI, and VARI exhibited comparable relationships with lentil desiccation responses. Conclusion Overall, integrating UAV-based multispectral imaging, vegetation indices, ML, and AUDPC provides a scalable, objective, and non-destructive framework for evaluating herbicide injury and desiccant efficacy under variable field conditions.