Leveraging Remote Sensing Data for Yield Prediction with Deep Transfer Learning

Area of Science:

• Agricultural Science
• Data Science
• Remote Sensing

Background:

• Automated yield prediction relies heavily on remote sensing data due to its resolution, availability, and cost-effectiveness.
• Challenges in yield prediction include obtaining reliable ground truth data and the expense of expert-acquired crop information.
• Deep transfer learning offers a solution by leveraging existing data patterns and reducing ground truth requirements.

Purpose of the Study:

• To develop and evaluate a deep transfer learning framework for crop yield prediction using remote sensing data.
• To transfer knowledge from US soybean yield prediction to Argentina, addressing data limitations.
• To enhance prediction accuracy by employing advanced transfer learning techniques and spatio-temporal analysis.

Main Methods:

• Utilized remote sensing data preprocessed into histograms for yield prediction.
• Implemented a deep transfer learning framework, including temporal alignment and techniques like L2-SP, BSS, and layer freezing.
• Applied Gaussian processes to exploit spatio-temporal patterns in the data.

Main Results:

• Achieved a 19% improvement in Root Mean Square Error (RMSE) and a 39% improvement in R-squared (R2) for soybean yield prediction in Argentina compared to baseline methods.
• Successfully transferred predictive knowledge from the US to Argentina, demonstrating the efficacy of the transfer learning approach.
• Validated the framework's ability to overcome challenges like catastrophic forgetting and negative transfer.

Conclusions:

• Deep transfer learning with remote sensing data provides a powerful tool for accurate yield prediction, particularly in regions with limited ground truth data.
• The proposed framework, incorporating advanced transfer learning techniques and Gaussian processes, significantly enhances prediction performance.
• This approach holds promise for improving agricultural monitoring and food security, especially in developing countries.