Abstract
We consider the problem of classifying a pair of Synthetic Aperture Radar (SAR) images by proposing an explainable and frugal algorithm that integrates a set of divergences. The approach relies on a statistical framework that takes standard probability distributions into account for modelling SAR data. Then, by learning a combination of parameterized Renyi divergences and their parameters from the data, we are able to classify the pair of images with fewer parameters than regular machine learning approaches while also allowing an interpretation of the results related to the priors used. Experiments on real multi-class data demonstrate the virtues of the suggested method when compared to both Random Forest and Convolutional Neural Networks (CNN) classifiers, showing its resilience to disturbances such as polluted labels and variations in the percentage of training data.
Introduction
Synthetic Aperture Radar (SAR) imaging is a cornerstone of modern Earth observation, providing high-resolution data regardless of weather or light conditions. However, SAR data is inherently affected by speckle noise—a multiplicative interference that complicates interpretation. While deep learning, particularly Convolutional Neural Networks (CNNs), has achieved high performance in image classification, it often requires massive datasets, significant computational power, and functions as a "black box," offering little insight into its decision-making process.
This research introduces a "frugal" and explainable alternative. By leveraging statistical physics and information theory, the proposed framework models SAR data through established probability distributions and learns optimal statistical distances—specifically Renyi divergences—to classify image pairs. This approach achieves competitive accuracy with a fraction of the parameters required by deep learning models.
Methodology: Frugal and Explainable Learning
The core of the proposed method is a shallow architecture designed to learn the optimal way to measure the "distance" between image patches.
Statistical Modeling
Rather than processing raw pixels, the framework models SAR data using three physically grounded distributions: Gamma, Log-normal, and Rayleigh. These distributions effectively capture how radar waves interact with different surface textures.
Renyi Divergence Learning
A divergence measures the distance between two probability distributions. The Renyi divergence is parameterized by $\alpha$, which dictates the sensitivity of the distance calculation. Unlike traditional methods that use a fixed $\alpha$ (e.g., $\alpha=0.5$ for the Bhattacharyya distance), this model treats $\alpha$ as a learnable parameter. This allows the model to discover the most effective "mathematical ruler" for each specific terrain type.
Architecture and Explainability
The resulting network is extremely lightweight, utilizing only 222 parameters compared to the 226,000+ parameters found in standard CNNs. By constraining the network weights to be positive, the authors ensure that the model remains interpretable; researchers can identify exactly which physical features and divergence types contributed to a specific classification.
Experimental Setup
The model was evaluated using an X-band SAR dataset from the PAZ satellite, focusing on the Mont Blanc massif. The dataset included 645 image patches (32x32 pixels) categorized into six classes: Glacier accumulation, City, Forest, Plain, Rocky areas, and a "Dissimilar" class for pairs representing different terrains. Ground truth labels were established using the CORINE land cover model.
Key Results and Discussion
Performance and Frugality
Despite its minimal parameter count, the Renyi model outperformed CNNs in identifying complex textures, such as Forests (80.5% vs 45.1%) and Cities (83.2% vs 61.1%). It also demonstrated superior resilience to limited training data, maintaining steady performance even with a 20% reduction in training samples, whereas CNN performance degraded significantly.
The Role of Alpha
Visualizing the learned $\alpha$ parameters provided unique insights into the classification process. While many $\alpha$ values clustered around 0.5 (the Bhattacharyya distance), they shifted toward 1.0 (Kullback-Leibler) for specific classes like forests and plains. This confirms that different physical terrains are best distinguished using different statistical metrics.
Robustness to Label Noise
The model's reliance on physical statistical priors makes it highly robust. In experiments where training labels were intentionally "polluted" or corrupted, the Renyi model maintained high accuracy, whereas purely data-driven models were more easily misled.
Conclusion
This research demonstrates that integrating domain-specific statistical priors into machine learning can produce models that are not only more efficient but also more transparent and robust than standard deep learning approaches. While the model currently struggles with the "Dissimilar" class—likely due to the assumption that image pairs follow the same distribution type—future work will focus on extending the framework to handle divergences between heterogeneous distributions and testing on more complex, unbalanced datasets.
Summary updated