Introduction
Aerial scene classification—the automated labeling of high-resolution satellite or drone imagery into categories like "airport," "stadium," or "forest"—is a fundamental task in remote sensing. However, standard deep learning architectures such as VGG-16 or AlexNet, designed for massive natural image datasets, often struggle with the "small data, big model" paradox in aerial contexts. These models contain tens of millions of parameters, leading to overfitting when trained on relatively small aerial datasets. To address this, researchers have developed JM-Net, a lightweight convolutional neural network (CNN), paired with a novel Cluster-SVM classification strategy.
Methodology
The proposed system follows a multi-stage pipeline designed for parameter efficiency and robust classification.
JM-Net Architecture
JM-Net (Justify Mentioned Net) is engineered to slash parameter counts while maintaining high feature extraction capability. Key innovations include:
- CEC Modules: The "Compress and Expand Convolution" module utilizes $1 \times 1$ filters to "compress" input channels before "expanding" them back using parallel $1 \times 1$ and $3 \times 3$ filters. This approach significantly reduces the model's footprint, as $1 \times 1$ filters have 9x fewer parameters than $3 \times 3$ filters.
- Removal of Fully Connected (FC) Layers: Traditional CNNs often house the majority of their parameters in final FC layers. JM-Net replaces these with a global average-pooling layer, extracting feature vectors without massive weight matrices.
- Delayed Downsampling: To preserve spatial details critical for aerial imagery, the authors postpone downsampling (pooling) until deeper layers.
Cluster-SVM Strategy
Rather than feeding raw feature vectors directly into a Support Vector Machine (SVM), the authors introduce a clustering step. Feature vectors are grouped into "piles," and the SVM classifies the centers of these clusters. This approach acts as a regularizer, reducing the likelihood of confusion between similar scenes. A custom similarity metric ($Sim\_loss$) is employed, combining Euclidean distance and cosine similarity to account for both the magnitude and direction of feature vectors.
Experimental Results
The model was evaluated on two major benchmarks: the Remote Sensing Dataset (RSD) and the Aerial Image Dataset (AID).
- Accuracy: JM-Net achieved 92.20% accuracy on RSD and 91.91% on AID, outperforming significantly larger models like VGG-16.
- Efficiency: The model achieves state-of-the-art results with only 1/40th the parameters of AlexNet, making it highly suitable for deployment on edge devices or satellites with limited storage and compute.
- Feature Representation: Ablation studies identified a 2048-dimensional feature vector as the "sweet spot" for aerial scene representation.
Conclusion and Future Work
JM-Net demonstrates that efficiency can drive accuracy in specialized domains like remote sensing. By reducing model noise and employing a clustering-based classification strategy, the system overcomes the limitations of traditional deep learning on small datasets. Future research may focus on reducing hyperparameter sensitivity in the Cluster-SVM stage and extending the JM-Net architecture to more complex tasks like object detection and semantic segmentation.
Summary updated