A Federated Learning-Based Crop Yield Prediction for Agricultural Production Risk Management

A spiralling global population and changing dietary needs have scaled up the demand for food and raw materials supplied to the industry. The agricultural production is struggling to keep up the level of required crop yields due to numerous risks affecting the yield. The past two decades have witnessed the increased intensity of agricultural production risks due to challenges posed by climate changes. There is a dire need to address it with proper insights into the data attributes impacting the crop yield. Currently, many of the machine learning and deep learning methods focus on training the model using the data collected and stored in a centralized data repository. However, many attributes related to weather data, soil data and crop management data are scattered and siloed to particular organization servers or smart farming devices. In this study, we are proposing a federated learning method for training yield prediction models on a horizontally distributed dataset located on different client devices. In particular, federated averaging algorithm is used to train the deep residual network based regression models such as ResNet-16 and ResNet-28 for soybean yield prediction in a decentralized setting and compare its performance with deep learning and machine learning methods. The results from experimented learning models show that federated averaging using ResNet-16 regression model with Adam optimizer yielded optimal results compared to centralized learning models and can be easily deployed for yield prediction in a federated setting.