Abstract

This study introduces a Convolutional Neural Network (CNN)-based approach for classifying cassava plant health using multispectral Unmanned Aerial Vehicle (UAV) imagery and vegetation indices. The vegetation indices, including Normalized Difference Vegetation Index (NDVI), Green Normalized Vegetation Index (GNDVI), EVI, Normalized Difference Red Edge (NDRE), and Leaf Chlorophyll Index (LCI), were statistically analyzed using one-way Analysis of Variance (ANOVA) and the Friedman test to evaluate their separability across different growth stages and health conditions. The CNN model was evaluated using datasets collected at 2, 4, and 7 months of cassava growth. The experimental results show that the proposed CNN achieved classification accuracies of 0.84, 0.81, and 0.96 for the 2-, 4-, and 7-month datasets, respectively, indicating that classification performance varied across growth stages. The lower accuracy at the intermediate stage was associated with increased canopy complexity and overlapping spectral characteristics. The proposed CNN was further compared with classical machine learning and pre-trained CNN baselines, including Random Forest, Support Vector Machine (SVM), XGBoost, ResNet50, and MobileNetV2. The proposed CNN demonstrated the most consistent performance across cassava growth stages, achieving the best classification results for the 2- and 7-month datasets, while XGBoost showed comparable performance for the 4-month dataset. The comparison before and after applying the Synthetic Minority Oversampling Technique (SMOTE) also indicates that class balancing improved CNN performance across all growth stages. Furthermore, Local Interpretable Model-Agnostic Explanations (LIME)-based interpretation shows that the model focused on relevant vegetation regions, thereby enhancing interpretability. Overall, the proposed framework demonstrates strong potential for UAV-based cassava health monitoring and provides insights into the relationship between spectral separability, class balancing, and classification performance.