Performance Analysis of CNN and Separable CNN for Land Use and Land Cover Classification in VHR Images Using Object-Oriented Approach with eXplainable Artificial Intelligence

Abstract

Land use and land cover (LULC) maps facilitate the generation of an accurate representation of the earth’s evolving natural ecosystem and the influences of human activities. Thanks to the improvements in satellite technology, the technique of imagine capture may now be conducted with a remarkable degree of accuracy and periodicity. Advanced deep learning algorithms have been developed to effectively handle and precisely analyze this data. This work aims to showcase the efficacy of the deep learning models utilizing the object-based approach with Convolutional Neural Network (CNN) and separable CNN architectures. Within the CNN architecture, every filter is applied to each input channel, facilitating the acquisition of more intricate characteristics. Additionally, it exhibits superior performance when dealing with extensive datasets and intricate jobs. The Separable CNN architecture offers advanced training and decreased memory consumption due to its reduced number of parameters and computational expenses. These models used WorldView-3 imagery to create LULC maps of Akyazı district, Turkey. In the beginning, segmentation was performed with Multiresolution segmentation for the segmentation stage. Then, spectral, texture and spatial features of each segment were calculated. A dataset containing train (70%), test (15%) and validation (15%) was created for training CNN and Separable CNN models separately. The CNN model demonstrated exceptional performance across all metrics. The Kappa accuracy of this model was 92.02% and the overall accuracy was 93.10%. In contrast, the Separable CNN model exhibited lower performance compared to the CNN model, with a Kappa accuracy of 90.68% and an overall accuracy of 91.95%. Future advancements and integration of both technologies have the potential to enhance performance and efficiency, particularly in large-scale and intricate LULC projects. This work is a significant advancement in expanding the possibilities and usefulness of both CNN and Separable CNN methods in object-oriented approach. © 2025 Elsevier B.V., All rights reserved.