Optimised multi-hierarchical feature fusion with multi-kernel CNN and spectral-spatial convolutions for remote sensing image classification

Abstract

Remote sensing image classification is a central process in the interpretation of Earth observation data for applications such as land use mapping, environmental monitoring, and urban planning. Despite significant progress brought about by deep learning, particularly Convolutional Neural Networks (CNNs), existing models are prone to failing to represent complex spatial patterns, multi-scale object variation, and rich spectral dependencies characteristic of high-resolution remote sensing images. In addition, most models are suffering from greater computational complexity and poor generalisation on diverse datasets. To address these issues, we propose an Optimised Multi-Hierarchical Feature Fusion Framework, a deep learning model that integrates multi-kernel convolution, spectral-spatial depthwise convolution, and residual learning into a ResNet-50 backbone. The model learns a rich set of spatial textures and spectral features at different hierarchical layers, leading to enhanced feature representation and classification robustness. We evaluated the proposed method on six remote sensing datasets: AID, EUROSAT, NWPU-RESISC45, RSSCN7, UC Merced, and WHU-RS19 with standard performance metrics like accuracy, precision, recall, specificity, and F1-score. The proposed method achieved an average accuracy of 99.00 % for AID, 99.51 % for EUROSAT, 99.56 % for RESISC45, 95.71 % for RSSCN7, 99.14 % for UC Merced, and 95.79 % for WHU-RS19. Moreover, qualitative visualisation techniques such as Class Activation Mapping (CAM) and LIME were employed to provide explanations of model decisions by identifying the most influential image regions to the predictions. These visual explanations confirmed that the model is based on semantically meaningful regions, further improving its interpretability and trustworthiness. The superior accuracy and robust performance on multiple datasets verify the efficiency and generalisation ability of the designed model, indicating it is a strong candidate for practical remote sensing classification tasks. https://github.com/chiagoziemchima/Multi-Hierarchical-Feature-Fusion-/tree/main.