Comparing parametric and nonparametric approaches for estimating forage availability using remote sensing and climatic datasets

Accurate estimation of forage availability is crucial for effective rangeland management. This study, conducted in a hot semi-arid rangeland (844.7 ha) in Khuzestan province, Iran, outlines a three-step approach to achieve this goal. Forage availability was measured using the cutting and weighing method on 58 one-square-meter sample plots. Initially, remote sensing (Sentinel-2) and climatic indicators were prepared, and the Pearson correlation coefficient was employed to analyze relationships between these independent variables and the dependent variable (forage mass). The second step focused on identifying optimal vegetation indices for forage availability estimation using multiple linear regression (MLR), Random Forest (RF), and classification and regression trees (C&RT) algorithms. The third step aimed to enhance model performance by integrating vegetation and climatic indices. Using 58 sample plots, 30 were used for modeling and 28 exclusively for final validation. A k-fold cross-validation (k = 5) was employed, with 47 plots for training and 11 for validation in each iteration. Model performance (MLR, RF, C&RT) was evaluated using average R² and RMSE metrics from all validation folds, ensuring robust and reliable results. The Random Forest used an internal 60–20-20 split. Correlation analysis revealed that MSAVI2 (-0.83), SAVI (-0.23), DVI (-0.17), potential evapotranspiration index (-0.42), and Transou index (0.35) were the most significant variables. In the second step, R² values for MLR, RF, and C&RT models were 0.54, 0.54, and 0.68, respectively. Incorporating potential evapotranspiration and Transou indices into the models boosted their coefficients of determination to 0.68, 0.72, and 0.91, with corresponding RMSE reductions of 85%, 69%, and 22%. The study concludes that a combination of the MSAVI2 index with potential evapotranspiration and Transou indices, implemented through the non-parametric C&RT regression, offers a robust method for estimating rangeland forage availability in hot, semi-arid, and arid environments.