Can Pixel-Based Approach Achieve Similar Performance to Area-Based Approach in Crop Yield Forecasting Using Sentinel-2 Imagery and Deep Neural Networks? A Probabilistic Analysis

A key solution to addressing food security challenges is the accurate forecasting of crop yields. Satellite imagery and historical remote sensing data play a crucial role in forecasting agricultural yields, particularly at the scale of individual fields. In the context of a small field, selecting between a pixel-based or area-based approach is essential, especially with deep learning (DL) models where having too little data can pose problems and acquiring more data can be difficult and challenging. In this study, we introduce two frameworks for forecasting corn yield: one using a pixel-based method and the other an area-based strategy, both utilizing Sentinel-2 imagery from a public image yield dataset and leveraging deep neural networks. The two methods were compared using key statistical metrics and also probabilistic techniques and analyses including error histogram of the strategies and their cumulative distribution function (CDF) plots considering the shapes of the lines, and the general distribution of the differences. Our findings underscored nuanced distinctions in the predictive capabilities, highlighting that the area-based model leaned towards greater accuracy in predictions, and the pixel-based approach emerged as a feasible substitute, delivering comparable performance despite slightly larger discrepancies.