Comparison of adaptive techniques to predict crop yield response under varying soil and land management conditions

Decision-making processes in agriculture often require reliable crop response models to assess the impact of specific land management. While process-based models are often preferred over empirical ones in current modelling communities, empirical crop growth models can play an important role in identifying the hidden structure of crop growth processes relating to a wide range of land management options. This study investigates the potential of predicting crop yield responses under varying soil and land management conditions by applying three different adaptive techniques: general linear models (GLMs), artificial neural networks (ANNs), and regression trees (RTs). The crop yield data used in this research consist of 720 maize yield indices from 11 different land management trials in southern Uganda. GLM showed the poorest results in terms of modelling accuracy, prediction accuracy, and model uncertainty, which might suggest its inability to model the non-linear causal relationships present in complex soil-land and crop-management interactions. The other two non-parametric adaptive models show significantly higher prediction accuracy than GLM. RT is the most robust technique for predicting crop yield at the study site. ANN is also a promising tool for predicting crop yield and offers insight into the causal relationships through the use of sensitivity analyses, but the complex parameterization and optimum model structure require further attention. The three adaptive techniques compared in this research showed different advantages and disadvantages. When these methods are used together, valuable information can be provided on crop responses, and more reliable crop growth models may result.