Use of Geostatistical Methods Involving Standard Deviation and Semivariogram to Assess the Accuracy of Water Surface Extraction by Supervised Classification

The extraction of water surfaces plays a pivotal role in various applications, yet it presents considerable challenges due to the intricate nature of homogeneous entities and the quality of satellite imagery. In this study, we evaluated the performance of four supervised classification algorithms, namely Minimum Distance (MD), Mahalanobis Distance (MHD), Maximum Likelihood (ML), and Support Vector Machine (SVM). Our investigation spanned four distinct study areas, encompassing two dams located in Eastern Algeria (Beniharoune and Ain Zada) and two wetland zones situated in the Western region (Lake of Dayet Oum Rhalez and Lake of Dhayat Morasli). To facilitate our assessment, we employed the Normalized Difference Vegetation Index (NDVI) to calculate standard deviation (SD) and semi-variogram values. The results indicated that the ML classifier consistently yielded favorable outcomes, with SD values of 0.027, 0.036, 0.037, and 0.007, while the MHD classifier proved less suitable for extracting homogeneous entities, yielding SD values of 0.137, 0.162, 0.036, and 0.122, respectively. This trend was also evident in the semi-variogram analysis, where the ML classifier demonstrated strong performance with values of 2.39e-004, 4.89e-007, 4.38e-004, and 1.52e-005, while the MHD classifier displayed contrasting results, with values of 9.46e-004, 5.18e-003, 4.46e-004, and 4.57e-004. Also, using data from the autumn season, the results consistently aligned with those obtained during the spring season, signifying the reliability of our approach across different timeframes. Our findings, in comparison to overall precision values, introduce a valuable approach for the assessment of classification accuracy in the context of homogeneous entities.