Machine learning-based cloud computing improved wheat yield simulation in arid regions

Combining machine learning (ML) with dynamic models is recommended by recent research for creating a hybrid approach for robust simulations but has received less attention thus far. Herein, we combined multi- ML algorithms with multi-crop models (CM_s) of the DSSAT platform to develop a hybrid approach for wheat yield simulation over 40 years in different locations. The simulation analysis included temperatures (minimum and maximum), solar radiation, and precipitation as important key ecological factors in wheat production that varied across sites and years. Detailed observed datasets of wheat yield from 1981 to 2020 were used for training and testing Artificial Neural Network (ANN), K-Nearest Neighbors (KNN), Random Forest Regressor (RFR), and Support Vector Regressor (SVR) using Google Colaboratory (Colab). Such models were built to create four main approaches, including two approaches as hybrid (CM_s-ML) and benchmark (pure ML), as well as two testing methods for each approach such as default (75 % training and 25 % testing) and warmest years (2001, 2006, 2009, 2010, and 2018). In addition to wheat yield simulations, ML approaches were used to identify the important features, improve accuracy, and reduce overfitting. We developed ML approaches by novel cells on the built models (i.e., pure ML and hybrid) to eliminate less important features from permutation. Our results revealed that ANN and RFR outperformed other ML algorithms (SVR and KNN) in wheat yield simulation accuracy. Application of ML algorithms reduced yield change from 31.7 % under DSSAT simulations to 8.1 % and uncertainty from 12.8 % to 7.2 % relative to observed wheat yield over the last four decades (1981–2020). Our novel approach, which includes a hybrid CM_s-ML model, cloud computing, and a new permutation tool, could be effectively used for robust crop yield simulation on a regional and global scale, contributing to better aid decision-making strategies.