Assessment of drought impact on agricultural production using remote sensing and machine learning techniques in Kairouan prefecture, Tunisia

Drought is a climate-induced disaster that can occur in all climatic regions with its features that varies considerably from one country to another. In addition, this natural disaster has wide-ranging impacts on water resources, ecosystems, energy, agriculture, forestry, human health and food security. Indeed, drought affects many economic activities and people. Its impacts are particularly serious in food-deficit countries with high dependence on subsistence agriculture. In Tunisia as many countries in MENA region, drought has a great effect on water problems in particular. It may conduct to increase the problem of water shortage and to affect agricultural productivities. For this reason, the Tunisian government encouraged and supported irrigated agriculture as a way to reduce the risk of drought. In this context, the improvement of water access and the irrigation for high value agricultural products were developed. Therefore, this work aims at predicting agricultural drought with the focus on its impact on olive growing farms using remote sensing data and advanced machine learning techniques. The approach developed in the current work is based on the use of multi-temporal medium resolution Landsat 8 OLI and TIRS imagery and the ensemble-based machine learning techniques for improving estimates of the drought indicators such as vegetation health index (VHI). The root-mean-square error (RMSE) and coefficient of determination (R2) and the cross-validation (CV) technique were employed to evaluate the performance of the proposed model. The capability of the proposed model was evaluated and compared with other machine learning algorithms, i.e., the random forests (RF), the support vector machine (SVM). The findings shows that the proposed model performed well (R² = 0.84, RMSE =1.38) and outperformed the remaining algorithms. It can be concluded that multi-temporal optical and thermal remote sensing combine with an advanced machine learning technique can be accurately used to estimate drought in semi-arid land areas. This approach may conduct to sustainable strategies of water resources in drought region.