Land Use Classification in Remote Sensing Images by Convolutional Neural Networks

Abstract: We explore the use of convolutional neural networks for the semantic classification of remote sensing scenes. Two recently proposed architectures, CaffeNet and GoogLeNet, are adopted, with three different learning modalities. Besides conventional training from scratch, we resort to pre-trained networks that are only fine-tuned on the target data, so as to avoid overfitting problems and reduce design time. Experiments on two remote sensing datasets, with markedly different characteristics, testify on the effectiveness and wide applicability of the proposed solution, which guarantees a significant performance improvement over all state-of-the-art references.

• The paper demonstrates that fine-tuning CNNs pre-trained on ImageNet boosts accuracy to 97.10% on the UC Merced dataset.
• It employs architectures like CaffeNet and GoogLeNet to tackle high intra-class variability and low inter-class distance in remote sensing imagery.
• The findings indicate that using CNNs significantly enhances the reliability of land use classification, promoting advanced geographic analysis.

Overview of "Land Use Classification in Remote Sensing Images by Convolutional Neural Networks"

The paper presented by Marco Castelluccio et al. explores the application of convolutional neural networks (CNNs) for semantic classification of remote sensing images. By leveraging architectures like CaffeNet and GoogLeNet, the authors explore various learning modalities: conventional training from scratch, utilization of pre-trained networks, and fine-tuning methods. Their work is focused on addressing the challenges of high intra-class variability and low inter-class distance, which complicate land use scene classification.

Methodology

The paper evaluates two remote sensing datasets—the UC Merced Land Use dataset and the Brazilian Coffee Scenes dataset—demonstrating the effectiveness of CNNs over previous approaches. The authors test different architectures and focus particularly on the advantages of fine-tuning CNNs pre-trained on the ImageNet dataset. This approach significantly boosts performance by utilizing the low-level features already learned in these networks, facilitating effective scene classification in both optical and non-optical images.

Results

The experimental results showcase a pronounced improvement over traditional methods. On the UC Merced dataset, the proposed CNN-based solutions achieve an accuracy of 97.10% with GoogLeNet, marking almost a 5% improvement over previous state-of-the-art methods. Similarly, for the Brazilian Coffee Scenes dataset, training GoogLeNet from scratch yields an accuracy of 91.83%, outstripping earlier results achieved via simpler methods like color descriptor-based BIC and SVM-classified CNN features.

Implications and Future Directions

The paper indicates that CNNs, particularly their fine-tuned variants, greatly enhance the accuracy and reliability of scene classification in remote sensing. This has implications for numerous applications in geographical analysis and environmental monitoring where accurate land use classification is vital. The findings suggest that as datasets continue to grow, the potential for training more complex and specialized models will further improve classification performance.

In theoretical terms, the research underlines the utility of pre-trained networks and the transferability of learned features across domains. It paves the way for future studies focusing on the adaptation of CNN architectures for even more diverse remote sensing modalities, such as SAR imagery, where the visible spectrum does not suffice and more sophisticated models are needed for accurate feature extraction.

Overall, this study provides a comprehensive comparison of CNN architectures and adaptation strategies, highlighting the relevance and applicability of deep learning techniques in remote sensing and land use classification. This research encourages continued exploration in adapting deep learning models to tackle challenges unique to remote sensing data.