Reduced dimensionality space of features using spectral indices for detecting changes in multitemporal Landsat-8 images

In this study, spatio-temporal changes in water covers are calculated for Lake Urmia (Iran) from 2013 to 2023 using multi-temporal Landsat-8 OLI (Operational Land Imager) and TIRS (Thermal InfraRed Sensor) multispectral images. The challenge was working in a reduced dimensionality space of characteristics. With this objective, ten spectral water indices have been assessed. Moreover, Land Surface Temperature (LST) maps have been obtained from water bodies for correlation purposes. The evaluation of the indices is developed by K-means (unsupervised) and Random Forest (supervised) machine learning non-parametric classifiers. The total agreement between the classification results and the test data in the case of the Salinity Water Index (SWI) exceeds 95 % (96.00–98.50 %) for Overall Accuracy (OA) and 0.80 (0.84–0.99) for F1-Score. Furthermore, a strong correlation is observed between the LST and the SWI index, with correlation coefficient values exceeding 0.77 in all cases. These results show the effectiveness of the SWI spectral index in detecting changes in water area, particularly for tracking variations over time in hyper-saline water bodies such as Lake Urmia. In this case, the water coverage observed from 2013 to 2023 shows a slight recovery in 2016, followed by a more significant increase in 2019 and 2020. However, in recent years, the lake has experienced a dramatic decline, with water coverage sharply decreasing by 2023. • Reduced dimensionality space of features simplifies the change detection processes. • Dimensionally reduced space from spectral index is an innovative strategy. • Results of the K-means classifier agree with those obtained using Random Forest. • Salinity Water Index is an optimal index for monitoring a saline lake (Urmia).