Decoding time: Unraveling the power of N-BEATS and N-HiTS vs. LSTM for accurate soil moisture prediction
Lisa Umutoni, Vidya Samadi, George Vellidis, Charles V. Privette, José O. Payero, Bulent Koç
Abstract
Deep neural networks (DNNs) can be trained to predict soil moisture dynamics, which is crucial for effective irrigation scheduling. However, a lack of interpretability in these networks constrains their efficacy in grasping the nuanced patterns prevalent in soil moisture time series data. This study is the first attempt known to the authors that develop interpretable DNNs to predict soil moisture fluctuations expressed as soil water tension across three root zone depths and prediction horizons. The Neural Hierarchical Interpolation for Time Series (N-HiTS) and Neural Basis Expansion Analysis Time Series (N-BEATS) models were used in this research. Historical soil water tension data collected at the University of Georgia’s C. M. Stripling Irrigation Research Park (SIRP) and in Blackville, South Carolina, were used to train and test the models. The results were benchmarked with the Long-Short-Term Memory (LSTM) to compare the models with a traditional, recurrent neural network. All the algorithms were coupled with a probabilistic multi-quantile loss function to quantify the uncertainty associated with predictions. Analysis suggested that the N-HiTS and N-BEATS models outperformed the LSTM across two testbeds by maintaining accuracy over the extended horizons and depths. The prediction uncertainty was more controlled for N-HiTS and N-BEATS with narrower uncertainty bands across horizons and soil depths, while LSTM exhibited widening intervals. We demonstrate how the proposed architecture can be augmented with uncertainty quantification to provide probabilistic soil water tension predictions that are interpretable without considerable loss in accuracy.