Crop yield prediction using a Hybrid Long Short-Term Memory-Gauss Newton optimization algorithm (LSTM-GNOA) model

Author(s): Krish Bansal


Solutions dedicated to individual, rural farmers that help predict crop yield are lacking as current solutions require specialised skills and are hidden behind paywalls essentially locking out secluded farmers from precision agriculture. A new approach has been explored in this paper where a long-short term memory network (LSTM) can accurately predict crop yield for 124 crops using a cost-efficient Gauss-Newton Optimization Algorithm to identify the global maxima of the coefficient of determination (R-squared). The mean square error (MSE) and R-squared value served as parameters to determine the model’s accuracy. To further validate the model’s accuracy, a range of other networks and optimization algorithms were deployed for comparison. The one-year-in-advance model demonstrated promising results with an average MSE of 0.21 and an average R- squared value of 0.82 for all 124 crops in the dataset.