Statistical and ANN-Based Modeling for Desertification Risk Prediction in Semi-Arid Regions: A Case Study of Nineveh Governorate

Abstract

Desertification and land degradation threaten food, water, and livelihood security across Iraq’s semi-arid north. We develop a station-based, hybrid framework for operational desertification-risk assessment in Nineveh Governorate using multi-decadal observations from Mosul, Tal Afar, and Rabiya. The approach couples (i) distribution-free trend diagnostics (Mann–Kendall with Sen’s slope and persistence control via trend-free pre-whitening and effective-sample-size corrections) with (ii) an interpretable composite Desertification Risk Index (DRI) built from directionally normalized indicators (temperature, humidity, wind, sunshine, rainfall) and objective CRITIC weights. For prediction, we evaluate horizon-explicit forecasts of DRI at 1, 3, 5, and 10 years using leakage-free rolling-origin splits, training-only transformations, and chronological refitting. Baselines (Persistence, Climatology) are compared with ARIMA, artificial neural networks (ANN), Random Forest (RF), and XGBoost using RMSE/MAE/R² and skill vs Persistence; pairwise differences are tested with Diebold–Mariano (squared-error, Newey–West). Across stations and horizons, most settings exhibit positive skill relative to Persistence. RF dominates at medium–long horizons—especially in Tal Afar and Rabiya (skill ≈0.70–0.82 at h=10)—while ANN is competitive at short leads (e.g., Tal Afar h=1 ≈0.36; Mosul h=3 ≈0.35). ARIMA is only competitive at Mosul )h=1(. However, no comparisons are statistically significant at p<0.05, reflecting small effective samples and high inter-annual variability. The framework provides reproducible diagnostics and horizon-aware outlooks that can augment persistence/climatology in early-warning workflows. Future extensions should incorporate remote-sensing predictors, homogenization, and probabilistic targets to strengthen both predictive utility and statistical confidence.