Field-scale UAV-based multispectral phenomics: Leveraging machine learning, explainable AI, and hybrid feature engineering for enhancements in potato phenotyping

Abstract

• We explore UAV-based multispectral imaging for phenotyping and disease detection in potatoes. • We auto-generate features with simple math and combine them with various vegetation indices. • This is the first study to integrate XGBoost, SHAP, and UMAP in potato phenotyping research. • We propose an analysis pipeline to enhance understanding of relevant agricultural traits. Fast and accurate identification of potato plant traits is essential for formulating effective cultivation strategies. The integration of spectral cameras on Unmanned Aerial Vehicles (UAVs) has demonstrated appealing potential, facilitating non-invasive investigations on a large scale by providing valuable features for construction of machine learning models. Nevertheless, interpreting these features, and those derived from them, remains a challenge, limiting confident utilization in real-world applications. In this study, the interpretability of machine learning models is addressed by employing SHAP (SHapley Additive exPlanations) and UMAP (Uniform Manifold Approximation and Projection) to better understand the modeling process. The XGBoost model was trained on a multispectral dataset of potato plants and evaluated on various tasks, i.e. variety classification, physiological measures estimation, and detection of early blight disease. To optimize its performance, nearly 100 vegetation indices and over 500 auto-generated features were utilized for training. The results indicate successful separation of plant varieties with up to 97.10% accuracy, estimation of physiological values with a maximum R 2 and rNRMSE of 0.57 and 0.129, respectively, and detection of early blight with an F1 score of 0.826. Furthermore, both UMAP and SHAP proved beneficial for comprehensive analysis. UMAP visual observations closely corresponded to computed metrics, enhancing confidence for variety differentiation. Concurrently, SHAP identified the most informative features – green, red edge, and NIR channels – for most tasks, aligning tightly with existing literature. This study highlights potential improvements in farming efficiency, crop yield, and sustainability, and promotes the development of interpretable machine learning models for remote sensing applications.