Vision Transformers and Convolutional Neural Networks for Land Use Scene Classification

Technical Summary: Vision Transformers and Convolutional Neural Networks for Land Use Scene Classification

Problem Statement
Land Use Scene Classification (LUSC) from remote sensing imagery is a fundamental task for environmental monitoring, urban planning, and resource management. While Convolutional Neural Networks (CNNs) have historically dominated this field due to their ability to capture local spatial features and their robustness in data-limited scenarios, they possess inherent limitations in modeling long-range spatial dependencies and global contextual relationships. Conversely, Vision Transformers (ViTs), which leverage self-attention mechanisms to process images as sequences of patches, offer superior capabilities for capturing global context but typically demand larger datasets and greater computational resources. The paper addresses the need for a systematic, controlled comparison of these two architectural paradigms to determine their respective strengths and suitability for diverse remote sensing applications.

Methodology
The study employs a rigorous comparative framework evaluating a representative CNN architecture (AlexNet) against a Vision Transformer architecture (ViT-B/16). To ensure a fair assessment, both models were subjected to identical experimental conditions, including:

• Datasets: Two benchmark remote sensing datasets were utilized: the UC Merced Land Use Dataset (2,100 high-resolution aerial images, 21 classes) and the EuroSAT Dataset (27,000 Sentinel-2 satellite images, 10 classes). Experiments were conducted on subsets containing 5 and 10 classes for each dataset.
• Data Partitioning: For all experiments, data was split into 80% training, 10% validation, and 10% testing.
• Preprocessing: Images underwent channel-wise standardization and stochastic augmentation (flipping, rotation, cropping, color perturbation) to reduce overfitting.
• Implementation: Both models were implemented using MATLAB on a system equipped with an AMD Ryzen 9 7950X processor and 128 GB RAM.
• Training Strategy: Both models utilized categorical cross-entropy loss, gradient-based optimization, early stopping, weight decay, dropout, and learning-rate scheduling.
• Evaluation Metrics: Performance was assessed using Accuracy, Precision, Recall, F1-score, and computational time.

Key Contributions and Results
The paper provides a direct performance comparison between CNNs and ViTs under consistent settings, yielding the following findings:

1. Performance on Limited Data (UC Merced): On the UC Merced dataset, which features approximately 100 samples per class, the AlexNet (CNN) model demonstrated superior performance compared to the ViT. The results suggest that CNNs are more effective when training samples are limited, likely due to their strong inductive biases (locality and translation invariance) which facilitate stable learning in data-scarce environments.
2. Performance on Large-Scale Data (EuroSAT): On the EuroSAT dataset, which contains significantly more samples per class (2,000–3,000), the ViT-B/16 model achieved slightly better classification accuracy than AlexNet. This indicates that ViTs can better leverage abundant data to learn global contextual representations and long-range dependencies that CNNs may miss.
3. Computational Efficiency: The study highlights a trade-off between performance and resource consumption. Training the ViT-B/16 model on the 10-class EuroSAT dataset required approximately 253 minutes, whereas AlexNet required only 137 minutes under the same hardware configuration. ViTs also exhibited higher memory consumption and computational complexity.

Significance and Claims
The paper concludes that neither architecture is universally superior; rather, the optimal choice depends on the specific constraints of the application.

• CNNs are identified as the robust choice for scenarios with limited training data or where local texture and spatial patterns are the primary discriminators.
• Vision Transformers are presented as the preferred option when sufficient training data is available and the task requires understanding complex global spatial relationships, despite the higher computational cost.

The study's significance lies in providing practical guidance for researchers and practitioners in remote sensing, offering evidence-based insights into selecting appropriate models based on dataset size and available computational resources. The authors do not claim to have solved the generalization problem across all domains but rather delineate the specific conditions under which each architecture excels.