Spatial Generalization of Crop Yield Prediction Models Using UAV Imagery and Machine Learning

Abstract

Global population is expected reach 9.7 Billions by 2050 leading to an increase in food demand significantly, placing then a pressure on agricultural systems to improve productivity in a sustainable manner. Precision agriculture, supported by remote sensing and machine learning (ML), has emerged as a promising tool for optimizing resource use and improving crop yield prediction. Among most popular remote sensing platforms, Unmanned Aerial Vehicles (UAVs) has been used widely due to its ability to capture high-resolution images. Despite encouraging results reported in prior studies, the ability of UAV-based ML models to generalize across different spatial contexts remains understudied. This thesis investigates the spatial generalization performance of ML models for crop yield prediction using UAV multispectral orthomosaics and yield monitor data collected from three agricultural fields - two canola fields and one corn field - located in Manitoba and Alberta, Canada. Linear Regression and Random Forest models are evaluated using a tile-based representation (5 m × 5 m) derived from five spectral bands and two vegetation indices, namely the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Red Edge (NDRE). Model performance were assessed under multiple training–testing configurations, including intra-field and inter-field baselines, Leave-One-Field-Out (LOFO) evaluation, and spatially aware sampling approaches such as Leave-One-Block-Out (LOBO) and Leave-One-Cluster-Out (LOCO). Results indicate that models trained on combined raw spectral bands consistently outperform those relying only on vegetation indices, suggesting that richer spectral information improves predictive performance. However, cross-field generalization remains limited, highlighting the challenges of transferring models across heterogeneous spatial environments. Performance variability between fields further emphasizes the importance of consistent data collection and evaluation protocols.