Machine Learning and Deep Learning Approaches for Multi-Class Land Cover Classification Using High-Resolution Satellite Imagery

Abstract

Land use and land cover (LULC) classification using Very High Resolution (VHR) satellite data is significant for a wide scope of applications in urban planning, environmental monitoring, engineering, and sustainable development. However, there is a lack of research on the performance comparison of different classifiers in optimal computing environments. In this research, twelve classification algorithms from Pixel-based, object-based image analysis (OBIA), patch-based deep learning, and graph and theory-based approaches are the four methodological paradigms that are looked at. In this study, we make a complete comparison between 12 classification algorithms belonging to four methodological paradigms for LULC classification using SPOT-5 multispectral imagery of the Iraqi city of Mosul. The methods that are being taken into consideration are Support Vector Machine (SVM), Random Forest (RF), Support Vector Machine (SVM) and Random Forest (RF) classifiers in SNIC-based OBIA, Seven models for deep learning (including CNN, ResNet, ViT, DaViT, U-Net with CNN, U-Net with ConvNeXt, and EfficientNet), and a Graph Convolutional Network (GCN). Our results demonstrate that while patch-based deep-learning techniques achieve higher accuracy, there is a significant difference between the accuracy of the various techniques. When using DaViT together with CNN, the best overall accuracy was reached, which is 99.83%, and has a kappa value of 0.9977. EfficientNet was in second place with 99.66%. The best performance was given by U-Net at 99.49%, followed by ViT at 99.41%. The classical pixel-wise RF showed an accuracy of 98.63% and can compete with much lower computational costs. The OBIA method with RF resulted in an accuracy of 98.67%, whereas the accuracy of the GCN-based method was only 78.16%. The McNemar's test result was very significant when comparing the best and worst performing models (p = 0.02). The analysis by class showed that the Road class was the hardest to classify in various approaches, while the Water class had a high level of accuracy. On average, Agriculture yielded the highest percentage of pixel misclassifications across all paradigms with only 12% pixels being correctly classified. Our findings provide empirical advice for selecting a method for VHR satellite image classification, taking into account the trade-offs among accuracy, computational efficiency, and practical implementation.