Field-Scale Rice Yield Prediction Using UAV Imagery and Machine Learning in a Developing Country Context

• DL model trained on MS data achieved the highest accuracy with an R 2 of 75.29% • Linear Regression Coefficient-Based feature selection with SVM and PCA with Linear Regression significantly improved model performance. • NIR and red-edge bands in the MS dataset consistently outperformed the RGB dataset • More represented rice variety (Sona) achieved a strong R 2 of 80.21% on MS data Accurate crop yield prediction is critical for agricultural planning, food security assessment, and farm-level decision-making. In Nepal, however, rice yield estimation is still predominantly based on traditional approaches, where local agricultural extension offices collect field-level observations that are subsequently aggregated at district, provincial, and national scales, often limiting spatial detail and timeliness. This study aims to develop a field-scale rice yield estimation framework by integrating Unmanned Aerial Vehicle (UAV)-derived remote sensing data with machine learning (ML) and deep learning (DL) techniques. High-resolution multispectral (MS) and RGB UAV imagery were used to evaluate the influence of Vegetation Indices (VIs), including HUE and VNDVI from RGB data and RGBVI and Simple Ratio (SR) from MS data, along with plant characteristics and farm management practices (e.g., application of Zyme and Zinc Potash) on rice yield. The predictive performance of Support Vector Machines (SVM), Linear Regression (LR), Decision Trees (DT), Random Forests (RF), and deep neural network models were systematically assessed. Data preprocessing included feature selection based on importance ranking, Yeo–Johnson power transformation, and Principal Component Analysis (PCA) to improve model stability and performance. Among conventional ML models, LR combined with PCA achieved a coefficient of determination (R²) of 69.09% using MS data, while SVM yielded the best performance using RGB data (R² = 68.27%). Overall, deep neural networks outperformed other models, achieving R² values of 75.29% and 64.60% for MS and RGB data, respectively. Model performance varied notably across rice varieties; the Sona variety (n = 127) achieved the highest coefficient of determination (R² = 80.21% for MS and 76.34% for RGB), whereas varieties with fewer samples exhibited lower predictive performance. Results further indicate that ranking features by importance, rather than eliminating them, enhances predictive accuracy, particularly when using LR-derived feature importance, which proved critical for improving the performance of both LR and SVM models.