Comparative Analysis of Deep Learning CNN Models and Traditional Machine Learning Approaches for Land Use Land Cover Classification Using Imagery

Abstract. An updated map of the area's land use land cover (LULC) is necessary for strategic planning and management of land use to shape the town sustainably. The advances in remote sensing imageries and artificial intelligence have facilitated the extraction of LULC classification. With the high number of studies on LULC mapping using various machine learning (ML) and deep learning (DL) algorithms incorporating imageries, no established algorithm shows stable results for all the datasets and study regions. Therefore, we used three robust machine learning algorithms, Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbour (KNN), and four deep learning algorithms, Residual Network (ResNet50 and ResNet152) and Visual Geometry Group (VGG16 and VGG19), to understand which model can produce a highly accurate LULC map in the Indian context, which are inherently unplanned and unorganized using Sentinel 2 imageries. The results of these models were then comparatively analyzed statistically using Accuracy, Recall, Precision, F1-score, and Kappa coefficient. Although DL models require a large number of training datasets, they outperformed the ML algorithms with higher Kappa coefficient values (ResNET50 = 0.90, ResNET-152 = 0.91, VGG-16 = 0.94, VGG-19 = 0.94). VGG-19 has consistently given better performance in all accuracy metrics. Overall the study highlights the potential of deep learning models, particularly VGG-19, in generating highly accurate LULC maps for complex and unplanned urban environments in India. These findings underscore the importance of leveraging advanced AI techniques in remote sensing for effective land use planning and sustainable urban development.