HDL: A Novel Hydrology Defined Loss Function for Enhanced Physics based deep learning for crop yield prediction

Abstract

Introduction-Agronomy is multi-dimensional and complex by nature. Agricultural prediction problems, especially yield prediction, exist in a universe of uncertainty since the highly non-linear and heterogeneous set of determinants exists. Although deep learning methods have been tried for yield prediction, their performance is constrained  by spatial granularity issues and not using enough domain knowledge.

Objectives-This paper responds to the questions raised earlier by highlighting two critical issues in making precise predictions of yield using deep learning models.

Method-The research provides a CNN-LSTM model for rice yield prediction with emphasis on block level (administrative unit of state ) spatial resolution for the Jammu district. While attempting to improve the accuracy of prediction, the research includes a physics-guided, hydrology-based loss function for the CNN-LSTM model.

Results-The results validate that the new loss function has made improved prediction compared to the baseline deep learning architectures of CNN, LSTM, and the basic CNN-LSTM.

Conclusion-The incorporation of a physics-informed, hydrology-sourced loss function into a CNN-LSTM model considerably alleviates issues of spatio-temporal variations  and lack of integration of domain knowledge, resulting in improved predictions of rice yield.