Adversarial attacks present significant threats to the robustness and security of neural networks, particularly in Remote Sensing Image (RSI) classification. The Fast Gradient Sign Method (FGSM) is a conventional technique for creating adversarial examples by perturbing input data along the gradient of the loss function. However, FGSM’s limitations, including its linearity assumption and fixed perturbation magnitude, often resulting in suboptimal adversarial examples. This paper presents an innovative approach to enhance FGSM by integrating it with Genetic Algorithm (GA) which optimizes the epsilon value iteratively to improve the attack's success rate and robustness. Experiments on EuroSAT and UCMerced-LandUse datasets utilizing deep learning architectures like ResNet, EfficientNet, and MobileNet demonstrate that the GA-enhanced FGSM significantly outperforms the baseline FGSM across several performance metrics, such as accuracy, precision, recall, F1 score, and loss. For instance, the MobileNet model’s accuracy dropped from 95.23% to 81.67% and 75% under FGSM and FGSM + GA attacks, respectively. Similarly, the performance of other models also declined significantly when subjected to the FGSM+GA attack. This integration of FGSM with GA not only boosts the efficacy of adversarial attacks but also offers insights into the underlying mechanisms of neural network vulnerabilities in remote sensing applications.