Artificial intelligence approach for the prediction of Robusta coffee yield using soil fertility properties

As a commodity for daily consumption, coffee plays a crucial role in the economy of several African, American and Asian countries; yet, the accurate prediction of coffee yield based on environmental, climatic and soil fertility conditions remains a challenge for agricultural system modellers. The ability of an Extreme Learning Machine (ELM) model to analyse soil fertility properties and to generate an accurate estimation of Robusta coffee yield was assessed in this study. The performance of 18 different ELM-based models with single and multiple combinations of the predictor variables based on the soil organic matter (SOM), available potassium, boron, sulphur, zinc, phosphorus, nitrogen, exchangeable calcium, magnesium, and pH, was evaluated. The ELM model’s performance was compared to that of existing predictive tools: Multiple Linear Regression (MLR) and Random Forest (RF). Individual model performance and inter-model performance comparisons were based on the root mean square error (RMSE), mean absolute error (MAE), Willmott’s Index (WI), Nash-Sutcliffe efficiency coefficient (ENS), and the Legates and McCabe’s Index (ELM) in the independent testing dataset. In the independent testing phase, an ELM model constructed with SOM, available potassium and available sulphur as predictor variables generated the most accurate coffee yield estimate (i.e., RMSE = 496.35 kg ha−1 or ±13.6%, and MAE = 326.40 kg ha−1 or ±7.9%). This contrasted with the less accurate MLR (RMSE = 1072.09 kg ha−1 and MAE = 797.60 kg ha−1) and RF (RMSE = 1087.35 kg ha−1 and MAE = 769.57 kg ha−1) model. Normalized metrics showed the ELM model’s ability to yield highly accurate results: WI = 0.9952, ENS = 0.406 and ELM = 0.431. In comparison to the MLR and RF models, the adoption of the ELM model as an improved class of artificial intelligence models for coffee yield prediction in smallholder farms in this study constitutes an original contribution to the agronomic sector, particularly with respect to the appropriate selection of most optimal soil properties that can be used in the prediction of optimal coffee yield. The potential utility of coupling artificial intelligence algorithms with biophysical-crop models (i.e., as a data-intelligent automation tool) in decision-support systems that implement precision agriculture, in an effort to improve yield in smallholder farms based on carefully screened soil fertility dataset was confirmed.