Enhancing Rice Yield Prediction using Vision Transformer and Regression Models

Amidst the growing global population, complex dynamics of climate change, and challenges in agricultural economies, ensuring food security and supporting agricultural policy planning have become more critical than ever. Thus, predicting rice yield before the harvest season is essential for billions who rely on rice as a staple food. In this study, we propose a novel approach for predicting rice yield using images of rice fields captured at the ripening stage. Our proposal leverages the Vision Transformer (ViT) architecture’s ability to extract relevant features from field images for yield prediction. Consequently, we replace the classification layer in the original ViT model with a regression layer to predict numerical values of rice yield. To determine the most effective regression method, we trained a variety of regression techniques. In an effort to develop a rice-specific prediction model for the Mekong Delta, we collected over 11,000 RGB images from 47 plots, encompassing more than 28 hectares in An Giang and Tra Vinh provinces, using digital cameras, smartphones, and fixed-wing UAVs. The rough grain yield recorded after harvest in these areas ranged from 6 to 12 tons per hectare (t/ha). Experimental results demonstrated that the ensemble regression model based on ViT achieved the lowest mean absolute error (MAE) of 72.72, underscoring the potential of ViT in agricultural yield forecasting.