Balancing centralisation and decentralisation in federated learning for Earth Observation-based agricultural...

Area of Science:

• Agricultural Science
• Remote Sensing
• Machine Learning

Background:

• Crop yield prediction using Earth Observation (EO) data faces challenges due to data diversity and limited access to proprietary datasets.
• Decentralised federated learning (FL) offers a privacy-preserving solution by avoiding direct data sharing, crucial for sensitive agricultural information.

Purpose of the Study:

• To investigate the impact of aggregation levels on the performance of federated learning for crop yield prediction.
• To analyze how aggregation interacts with other parameters to generalize findings for diverse crop yield datasets.

Main Methods:

• A proxy model was trained using crop type data from Copernicus Sentinel-2 images.
• Federated learning aggregation levels were systematically simulated and analyzed.
• Current and future crop yield dataset distributions were examined to identify optimal aggregation strategies.

Main Results:

• The study found that aggregation levels significantly influence federated learning performance in crop yield prediction.
• Dataset size was identified as a critical factor affecting both learning outcomes and the level of privacy preserved.
• Interactions between aggregation levels and other parameters were simulated to understand generalizability.

Conclusions:

• Optimal aggregation levels are crucial for effective decentralized federated learning in agriculture.
• Dataset size is a key determinant for balancing model accuracy and data privacy in federated crop yield prediction.
• Findings provide insights for designing future decentralized federated learning architectures for crop yield estimation.