Vision Transformers and Convolutional Neural Networks for Land Use Scene Classification

A new comparative study evaluates Vision Transformers against Convolutional Neural Networks for land use scene classification in remote sensing imagery, finding that each architecture holds distinct advantages depending on data availability and scene complexity [1]. The paper, submitted to arXiv on 20 May 2026 and revised on 2 Jun 2026, was authored by Arun Kulkarni [1]. It benchmarks a representative CNN, AlexNet, against a Vision Transformer using the UC Merced Land Use and EuroSAT Land Use datasets [1]. The study examines classification accuracy, precision, recall, F1-score, and computational complexity [1]. Experimental results indicate that CNNs perform robustly on datasets with limited training samples and strong local texture characteristics [1]. In contrast, Vision Transformers exhibit superior performance in capturing global spatial relationships in complex scenes when sufficient training data are available [1]. The paper notes that ViTs typically require greater computational resources and larger training datasets to achieve optimal performance [1]. This trade-off is consistent with broader trends in remote sensing. A separate 2024 study comparing CNN and transformer-based methods on the EuroSAT dataset also found that transformer-based models outperformed CNNs in some cases, achieving state-of-the-art results [4]. That study trained ten models—seven convolutional architectures including AlexNet, ResNet50, and ConvNeXt, and three transformer-based architectures including ViT and Swin Transformer—and reported that both model families achieved high accuracy [4]. Further research into on-board satellite processing has reinforced the viability of transformer models under operational constraints. A comparative study of pre-trained vision transformer models for Earth observation found that EfficientViT-M2 achieved 98.76% accuracy, precision, and recall, while demonstrating greater robustness under noisy inference conditions with an overall robustness score of 0.79 [5]. MobileViTV2 excelled on clean validation data, but EfficientViT-M2 proved more resilient to the noisy conditions common in satellite-based inference [5]. Land use scene classification is a critical component of environmental monitoring, urban planning, and sustainable resource management [1]. The findings from Kulkarni's study provide guidance for selecting appropriate models, suggesting that the choice between CNNs and ViTs should be driven by dataset size, computational budget, and the spatial complexity of the target scenes [1].