Reproduce “Simultaneous Corn and Soybean Yield Prediction from Remote Sensing Data Using Deep Transfer Learning” on OpenResearch

About this paper

Saeed Khaki, Hieu Pham, Lizhi Wang

Researchers from Iowa State University and Syngenta Seeds introduced YieldNet, a deep learning framework for simultaneous corn and soybean yield prediction using remote sensing data. This approach leverages a shared convolutional backbone and a novel loss function, achieving Mean Absolute Errors of 7.8% for corn and 7.67% for soybean by October predictions, while significantly reducing computational time and model parameters compared to single-crop models.

Key insights

• Simultaneously predicting multiple crop yields using a shared deep learning backbone (transfer learning) enhances individual crop prediction accuracy and computational efficiency by leveraging transferable information.
• Converting multi-spectral time-series images into 3-D histogram representations effectively captures essential spectral and temporal information for yield prediction, circumventing challenges with raw pixel data for deep transfer learning.
• A custom 'max' normalized loss function is crucial for multi-target regression problems to ensure all outputs are effectively optimized, leading to balanced and improved overall performance.

Summary by alphaXiv.