Representation Learning with a Variational Autoencoder for Predicting Nitrogen Requirement in Rice

Abstract

The scope of this research was to provide rice growers with optimal N-rate recommendations through precision agriculture applications. To achieve this goal, a prediction rice yield model was constructed, based on soil data, remote sensing data (optical and radar), climatic data, and farming practices. The dataset was collected from a rice crop surface of 89.2 ha cultivated continuously for a 5-year period and was analyzed with machine learning (ML) systems. A variational autoencoder (VAE) for reconstructing the input data of the prediction model was applied, resulting in MAE of 0.6 tn/ha, with an average yield for the study fields and period measured at 9.6 tn/ha. VAE learns the original input data representation and transforms them in a latent feature space, so that the anomalies and the discrepancies of the data are reduced. The reconstructed data by VAE provided a more sophisticated and detailed ML model, improving our knowledge about the various correlations between soil, N management parameters, and yield. Both optical and radar imagery and the climatic data were found to be of high importance for the model, as indicated by the application of XAI (explainable artificial intelligence) techniques. The new model was applied in the 2022 rice cultivation in the study fields, resulting in an average yield increase of 4.32% compared to the 5 previous years of experimentation.