Evaluation and uncertainty assessment of wheat yield prediction by multilayer perceptron model with bayesian and copula bayesian approaches

An accurate crop yield forecast is vital for sustaining food security and preventing famine. Artificial intelligence provides a robust tool for integrating environmental, physiological, and management data and predicting crop yield. In this study, we used a Multilayer Perceptron (MLP) model in its default and hybrid learning modes to predict wheat yield. In its default mode, MLP was trained with the backpropagation algorithm, and in the hybrid mode, MLP was trained with the Water Striders Algorithm (WSA), Sine-Cosine Algorithm (SCA), and Genetic Algorithm (GA). We further considered two ensemble modeling approaches, including the Bayesian Model Averaging (BMA) and Copula-based Bayesian Model Averaging (CBMA). The study area was classified into four homogenous wheat cultivation regions using the Fuzzy clustering approach, and for each region, the models were trained based on 30 years of fertilizer information, harvested yield, and climatic data. Our results showed that hybridizing MLP with an optimization algorithm provides more accurate yield predictions than its default mode. WSA required the lowest computational time, and its combination with MLP resulted in a more accurate yield prediction compared to other optimization algorithms. The accuracy of yield predictions was further improved when ensemble modeling was implemented. Accordingly, the CBMA approach generated the most accurate result with the lowest uncertainty range. The mean absolute error (MAE) of the CBMA was 0.026, 0.058, 0.088, 0.133, and 0.0201 lower than that of the BMA, MLP-WSA, MLP-SCA, MLP-GA, and MLP models. GLUE uncertainty analysis verified the superiority of the hybrid models and ensemble approaches. The predicted yield by CBMA, BMA, and MLP-WSA covered 94 %, 92 %, and 89 % of the observations, respectively.