A comparison of GEOBIA Vs PBIA machine learning methods for lithological mapping using Sentinel 2 imagery: Case study of Skhour Rehamna, Morocco

An extensive range of remote sensing-based approaches has been used to lithological mapping in conjunction with multispectral satellite imagery. Thus, this study focuses on mapping and comparing thematic classification methods of lithologies using Sentinel 2 imagery in the southern of Skhour Rehamna. First, Pixel-Based Image analysis (PBIA) techniques were evaluated, particularly, RF and K-NN machine learning algorithms (MLAs), both methods generalized lithological maps that is characterized with “salt and pepper” noise of misclassified pixels and poorly defined boundaries. The second approach is Geographic ObjectBased Image Analysis (GEOBIA), which involved multiresolution segmentation part based on several predefined scale, shape and compactness parameters that must be subjected to relative weighting in order segment the image to homogeneous areas, whilst, standard Nearest Neighbor classification was applied to the given samples. Then, classifications accuracies were assessed by comparing both methods using confusion matrix and Kappa coefficient. The results show that the GEOBIA approach provides the best mapping results with an overall accuracy of approximately 83.46% and Kappa coefficient of 0.76 compared with the results obtained by PBIA approach, with 81.92% and 80.79%, and Kappa coefficient of 0.72 and 0.72, for RF and K-NN respectively. As a conclusion, the results clearly indicate the potential of GEOBIA technique for lithological mapping.