Land Cover Classification Based on Machine Learning and Deep Learning Methods Using Sentinel-2 Satellite Images: A Case Study of the Urban Area in West Tehran

The production of land cover maps (LCM) provides essential information about land types and their characteristics, playing a significant role in updating urban maps, managing natural resources, environmental protection, and sustainable development. In this context, the use of image processing techniques and free remote sensing data is considered an optimal method for generating land cover maps (LCM). In this study, various artificial intelligence approaches, including machine learning (ML) and deep learning (DL) algorithms, were used to produce the LCM. The ML approach includes two stages: feature extraction and classification. In the feature extraction stage, texture features extracted from the gray-level co-occurrence matrix (GLCM), including mean, variance, homogeneity, contrast, and entropy, were used. For classification, the logistic regression (LR), decision tree (DT), random forest (RF), and support vector machine (SVM) algorithms were employed. In the DL approach, deep learning semantic segmentation models, including U-Net, U-Net++, ResU-Net, and MRU-Net, were used. To evaluate the accuracy of the ML and DL algorithms in producing the land cover map, Sentinel-2 images from two areas located in the west of Tehran were utilized. The results of this study were examined in three different sections: ML, DL, and their comparison. In the ML section, the RF model, which used a combination of the image's primary bands and texture features, performed better than other models with an overall accuracy of 95.21% and a Kappa coefficient of 92.62%. In the DL section, the MRU-Net model produced the most optimal LCM with an overall accuracy of 95.33% and a Kappa coefficient of 92.73% compared to other deep models. The MRU-Net model, without using texture features, improved overall accuracy and the Kappa coefficient by 0.53% and 0.82%, respectively, compared to the RF model using a combination of primary image bands. Furthermore, compared to the RF model, which used a combination of primary bands and texture features, the MRU-Net model's overall accuracy and Kappa coefficient were 0.12% and 0.11% higher, respectively.