TEGOA-CNN: An Improved Gannet Optimization Algorithm for CNN Hyperparameter Optimization in Remote Sensing Sence Classification

The evolution of remote sensing technology has led to significant improvements in high-resolution and hyperspectral image acquisition, enhancing applications like environmental monitoring and disaster assessment. However, the high dimensionality, nonlinearity, and heterogeneity of these images pose challenges for intelligent interpretation. While deep learning models (e.g., CNN) require balancing efficiency and parameter optimization, meta-heuristic algorithms establish self-organizing, parallelized search mechanisms capable of achieving asymptotic approximation towards the global optimum of parameters without requiring gradient information. In this paper, we first propose an improved Gannet Optimization Algorithm (GOA), named TEGOA, which uses the T-distribution perturbation and elite retention to address CNN’s parameter dependency. The experiment on CEC2017 shows that TEGOA has a better performance on composition functions. Hence, it is suitable for solving complex optimization problems. Then, we propose a classification model TEGOA-CNN, which combines TEGOA with CNN to increase the accuracy and efficiency of remote sensing sence classification. The experiments of TEGOA-CNN on two well-known datasets, UCM and AID, showed a higher performance in classification accuracy of remote sensing images. Particularly, TEGOA-CNN achieves 100% classification accuracy on 10 out of the 21 surface categories of UCM.