Early Wheat Yield Prediction Using Sentinel-1 Data: A 2D-CNN Deep Learning Approach with Machine Learning Integration

Early and accurate yield prediction is essential for ensuring food security and optimizing resource allocation in agriculture. Traditional yield prediction models often rely on either comprehensive crop cycle data or simplified single-stage data, leaving a gap in understanding the optimal data collection timing. This paper addresses this gap by critically assessing three key factors: model selection, input features, and the optimal crop stage for data collection. We evaluate various machine learning (ML) and deep learning (DL) models to determine the most effective approach for yield prediction. Additionally, we explore the impact of different input features and identify the crucial crop stages necessary for accurate predictions. To overcome the limitations of multispectral data, which is often affected by cloud coverage, we utilize Sentinel-1 synthetic aperture radar (SAR) data. SAR’s cloud-independent capabilities make it ideal for consistent time-series analysis. Our findings aim to enhance yield prediction accuracy by identifying the most informative crop stages and leveraging SAR data, thereby contributing to more efficient agricultural planning and resource management.