Accurate land use and land cover (LULC) classification remains challenging when visually and spectrally similar classes coexist, particularly in RGB-based remote sensing imagery. While numerous machine learning (ML) and deep learning (DL) models report high overall accuracy, limited attention has been given to the spatial structure and persistence of misclassification errors. In order to gain more insight into model-dependent and systematic error patterns, this study introduces a misclassification-driven evaluation perspective through a controlled cross-architecture comparison of custom Convolutional Neural Networks (CNNs), AlexNet, and ResNet50 models applied to the EuroSAT dataset under a fixed RGB-only setting. This approach reduces input-related variation and allows a controlled comparison of model behavior across architectures. We demonstrate that models with similar overall accuracy can exhibit different class-wise error distributions. Across the evaluated architectures, misclassification patterns were stable for specific confusable class pairs (notably Annual Crop vs Permanent Crop), indicating that RGB-only inputs impose a dominant limitation that persists across model families. These results show that improving the reliability of land-use and land-cover classification requires more than changing models; it also requires richer feature representation and greater data diversity. The paper identifies several shortcomings that must guide the next generation of multispectral and domain-adaptive research and provides practical suggestions regarding the improvement of accuracy and robustness of LULC classification systems. • This study compared CNN, AlexNet, and ResNet50 for land use and land cover classification tasks. • Models were trained using RGB EuroSAT images with transfer learning applied. • ResNet50 variant 3 offered the best result with 96.96% test accuracy. • Major misclassifications occurred between visually similar crop categories. • Future work should include multispectral data, hybrid models, and XAI tools.
Alam et al. (Wed,) studied this question.