Drought Detection in Satellite Imagery: A Layered Ensemble Machine Learning Approach

Drought has been a major calamity due to climate change in recent years. Predicting drought has grabbed the attention of meteorologists and climate scientists, who study and look for modern techniques. The early detection of drought leads to better management of resources and timely decisions to avoid damage. Machine learning techniques have proven their potential in classification and prediction problems. This study proposes a layered ensemble machine learning approach to detect drought from satellite imagery. The satellite imagery is collected using Google Earth Pro for the region ’Tharparkar’ in Pakistan’s Sindh province. Tharparkar is one of the most drought-stricken regions in Pakistan. The proposed approach combines conventional machine learning algorithms (Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Naive Bayes (NB), and k-Nearest Neighbor (k-NN)) with ensemble methods (Bagging and Voting) in a layered fashion for detecting drought from satellite imagery. The novelty of the study lies in its layered ensemble architecture that integrates multiple conventional classifiers with ensemble techniques for improved drought detection accuracy using satellite imagery. The validation of the model is computed using the stratified split method. The developed classification model is evaluated using well-established indexes, including accuracy, precision, recall, F measure, and Area Under Curve (AUC). Among the classical models, the Decision Tree classifier performed best with an accuracy of 82.17%, a precision of 82.53%, a recall of 82. 28%, and an F1 score of 82.35%. Within the Bagging models, Bagged Decision Tree achieved the highest performance, attaining an accuracy of 84.78%, a precision of 85.14%, a recall of 84.87%, and an F1 score of 84.91%. The final-layer Voting ensemble outperformed all previous models, yielding the highest F1 score of 84.80%. Based on experimental results, the proposed model has strong potential for practical deployment in real-world environmental monitoring systems.