Advanced Yield Prediction in Rice Using UAV Multispectral Imaging and Machine Learning Approaches in Sri Lanka

Abstract

Rice is an essential food crop making a significant contribution towards food security worldwide, particularly in Asian countries like Sri Lanka where it is widely grown. Accurate prediction of rice yields is critical for enhancing resource management and minimizing agricultural risks. Traditional yield estimation methods are often labor-intensive and prone to inaccuracies. However, advancements in technology, particularly remote sensing, offer a promising alternative. The aim of this study is to predict paddy yields three commonly grown rice varieties in Sri Lanka (Bg 352, At 362, and Bg 360) non-destructively using machine learning and multispectral images. Information gathered from UAV-based multispectral images, such as the NDVI, LCI, and EVI vegetative indices, were used to build models for predicting outcomes. The performance of Random Forest (RF), Support Vector Regression (SVR), and Simple Linear Regression (LR) in predicting rice yield was compared. The results indicated that the Random Forest Model trained with NDVI data extracted from UAV images captured at booting stage, showed the higher precision with an R² of 0.81, 0.76, 0.69 for Bg 352, BG 360 and AT 362, respectively. These results could be effectively used to provide timely and accurate yield forecasts for farmers and help them to allocate resources better and improve productivity. Future research based on such precision agriculture technologies should explore additional environmental variables and expand model validation across different growing seasons and regions.