Accurate mapping of land use and land cover (LULC) is essential for monitoring environmental changes, particularly in mining-impacted semi-arid landscapes where anthropogenic disturbances significantly alter ecosystem structure and function. This study assesses the performance of six machine learning classifiers Random Forest (RF), Gradient Tree Boost (GTB), Support Vector Machine (SVM), k-Nearest Neighbours (KNN), Minimum Distance (MinDist), Classification and Regression Tree (CART), and three ensemble strategies (Stacked, Hybrid Stacked, and Weighted Voting) for LULC classification using Landsat 7 and 8 imageries in the Acland region, Queensland, Australia between 2000 and 2024. Spectral bands and derived indices (NDVI, NDWI, NDBI, SAVI, EVI, BSI) were used as input features within a Google Earth Engine framework. The results indicate that RF and GTB consistently outperformed other base classifiers, while SVM showed a substantial improvement with higher-quality Landsat 8 data. Ensemble strategies, particularly hybrid stacked ensembles, further enhanced accuracy, producing smoother thematic maps, and improving separability in spectrally overlapping classes such as bare soil, mining areas, and light vegetation. Sensitivity analysis showed that classifier performance is affected by the initialization of pseudo-random number generators (seeds). Hybrid ensembles exhibited reduced variability and increased robustness between different seeds, outperforming standard stacking. Using identical seeds for both base and ensemble classifiers reduced diversity and degraded performance, underscoring the critical importance of randomization in ensemble learning. Comparisons between Landsat 7 and 8 highlight the benefits of improved radiometric quality, which reduces variability and enhances classifier stability. Overall, this study demonstrates that ensemble classification strategies provide a robust approach to long-term LULC monitoring in heterogeneous, mining-disturbed landscapes with the ability to integrate complementary classifier strengths, accommodate sensor differences, and maintain stability across random initializations.