Influence of Parameters in SDM Application on Citrus Presence in Mediterranean Area

Abstract

Within the context of Agriculture 4.0, the importance of predicting species distribution is increasing due to climatic change. The use of predictive species distribution models represents an essential tool for land planning and resource conservation. However, studies in the literature on Suitability Distribution Models (SDMs) under specific conditions are required to optimize the model accuracy in a specific context through map inspection and sensitivity analyses. The aim of this study was to optimize the simulation of the citrus distribution probability in a Mediterranean area based on presence data and a random background sample, in relation to several predictors. It was hypothesized that different parameter settings affected the SDM. The objectives were to compare different parameter settings and assess the effect of the number of input points related to species presence. Simulation of citrus occurrence was based on five algorithms: Boosted Regression Tree (BRT), Generalized Linear Model (GLM), Multivariate Adaptive Regression Splines (MARS), Maximum Entropy (MaxEnt), and Random Forest (RF). The predictors were categorized based on 19 bioclimatic variables, terrain elevation (represented by a Digital Terrain Model), soil physical properties, and irrigation. Sensitivity analysis was carried out by (a) modifying the values of the main models’ parameters; and (b) reducing the input presence points. Fine-tuning the parameters for each model according to the literature in the field produced variations in the selection of predictors. Consequently, probability changed in the maps and values of the accuracy measures modified. Results obtained by using refined parameters showed a reduced overfitting for BRT, yet associated with a decrease in the AUC value from 0.91 to 0.81; minor variations in AUC for GLM (equal to about 0.85) and MARS (about 0.83); a slight AUC reduction for MaxEnt (from 0.86 to 0.85); a slight AUC increase for RF (from 0.88 to 0.89). The reduction in presence points produced a decrease in the surface area for citrus probability of presence in all the models. Therefore, for the case study analyzed, it is suggested to keep input presence points above 250. In these simulations, we also analyzed which covariates and related ranges contributed most to the predicted value of citrus presence, for this case study, for different amounts of input presence points. In RF simulations, for 250 points, isothermality was one of the major predictors of citrus probability of presence (up to 0.8), while at increasing of the input points the contribution of the covariates was more uniform (0.4–0.6) in their range of variation.