💻 computer science

Feature-Space Oversampling for Addressing Class Imbalance in SAR Ship Classification

This paper proposes two novel feature-space oversampling algorithms, M2m $_f$ and M2m $_u$ , which effectively address class imbalance in SAR ship classification by significantly improving F1-scores on OpenSARShip and FuSARShip datasets compared to existing methods.

Original authors: Ch Muhammad Awais, Marco Reggiannini, Davide Moroni, Oktay Karakus

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ch Muhammad Awais, Marco Reggiannini, Davide Moroni, Oktay Karakus

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Unfair Classroom"

Imagine you are a teacher trying to teach a class about different types of ships. You have a huge textbook (the dataset) with pictures of ships. However, the book is unbalanced:

There are 1,000 pictures of big cargo ships.
There are only 50 pictures of fishing boats.

If you train a student (an AI model) on this book, they will become an expert at spotting cargo ships but will fail miserably at identifying fishing boats. They will just guess "cargo ship" every time because that's what they see most often. In the world of radar (SAR) images, this is a huge problem because these images are often grainy and noisy, making it hard to teach the AI with just a few examples.

The Old Solution: "Copy and Paste" (and why it failed)

Usually, when teachers don't have enough examples of the rare items, they try to "augment" the data. This is like taking a single photo of a fishing boat, flipping it upside down, making it slightly brighter, or zooming in, and pretending these are new, unique photos.

The authors say this doesn't work well for radar images. Because radar images are so noisy and low-resolution, simply copying and pasting them creates fake-looking images that confuse the AI rather than helping it.

The New Solution: "The Feature Space Bakery"

Instead of trying to make more pictures of the rare ships, the authors decided to work with the essence of the ships.

Imagine the AI has already looked at all the pictures and written a "flavor profile" for each one.

Cargo Ship Flavor: "Heavy, boxy, long."
Fishing Boat Flavor: "Small, agile, specific shape."

The authors propose a new method called Feature-Space Oversampling. Instead of baking new cakes (images), they take the "flavor profiles" (mathematical features) of the common ships and mix them with the rare ships to create synthetic flavor profiles.

They created two new recipes (algorithms) to do this mixing:

Recipe M2mf (The "Safe Mixer"):
- This method takes a "flavor profile" from a common ship and gently nudges it toward the "center" of the rare ship's flavor.
- The Safety Check: Before adding this new synthetic flavor to the mix, it checks: "Is this new flavor too similar to the ones we already have?" If it's too similar, it throws it away. This ensures every new example is unique and adds something new to the learning process.
Recipe M2mu (The "Similarity Matcher"):
- This method also nudges common flavors toward the rare ones.
- The Safety Check: Instead of checking for uniqueness, it checks: "Does this new flavor actually look like a real rare ship?" It uses a "similarity score" to make sure the synthetic example isn't drifting too far away from what a real fishing boat actually looks like.

The Experiment: The Taste Test

The authors tested these two recipes on two different "cookbooks" (datasets):

OpenSARShip: A standard collection of radar ship images.
FuSARShip: A newer, higher-quality collection.

They used three different "taste testers" (AI models: ViT, VGG16, and ResNet50) to see which recipe worked best.

The Results: What Worked?

The Old Way Failed: The traditional method of copying and pasting images (called M2morig) actually made the AI perform worse or stayed the same. It confused the AI.
The New Way Succeeded: Both of the authors' new "flavor mixing" recipes (M2mf and M2mu) significantly improved the AI's ability to spot the rare ships.
- On the FuSARShip dataset, the M2mu recipe was the winner, boosting the AI's accuracy by nearly 9%.
- On the OpenSARShip dataset, the M2mf recipe was slightly better, boosting accuracy by about 5%.

The "Underfitting" Surprise:
One of the taste testers (ResNet50) struggled in almost every scenario. The authors suggest this model was too "deep" and complex for the limited amount of data, causing it to get confused (underfit), whereas the other models (ViT and VGG) handled the new recipes much better.

The Bottom Line

The paper proves that for radar ship classification, it's better to mix the "essence" of the data than to copy the pictures.

By creating synthetic features (the mathematical description of the ship) rather than synthetic images, the AI learns to recognize rare ships much better. This approach is also cheaper and faster because you don't have to re-process the entire image; you just mix the mathematical "flavors" and train the final classifier.

In short: Don't try to fake more photos of rare ships. Instead, mathematically blend the "vibe" of common ships to create new, realistic examples of the rare ones, and your AI will learn much faster.

Technical Summary: Feature-Space Oversampling for Addressing Class Imbalance in SAR Ship Classification

Problem Statement
Synthetic Aperture Radar (SAR) ship classification is critical for maritime traffic monitoring but faces significant challenges due to data scarcity and severe class imbalance. Traditional deep learning approaches often struggle with these long-tailed datasets. While data augmentation in the sample space (image level) is common for optical data, it frequently fails for SAR imagery due to inherent noise and low resolution. Furthermore, standard transfer learning, which freezes feature extractors, does not inherently solve the imbalance issue. The authors identify a need for methods that can generate synthetic data to balance classes without relying on fragile image-level manipulations.

Methodology
The proposed framework shifts the oversampling process from the sample space to the feature space. The pipeline consists of three stages:

Feature Extraction: A pre-trained model (VGG16, ResNet50, or ViT) with frozen weights extracts features from SAR images.
Feature Oversampling: Two novel algorithms, M2mf and M2mu, are applied to the extracted features to generate synthetic samples for minority classes using features from majority classes.
- M2mf (Major-to-minor feature): Inspired by the original sample-space M2m method, this algorithm calculates class centroids for minority classes. It generates synthetic features by perturbing majority class features toward these centroids. A synthetic feature is retained only if it maintains a minimum distance ( $d_t = 0.5$ ) from existing synthetic features to ensure uniqueness.
- M2mu (Major-to-minor unique): This variation retains a synthetic feature only if it is "similar" to the original features of the target minority class. Similarity is measured using Cosine similarity, and a feature is kept only if it exceeds a threshold ( $sim_t = 0.8$ ).
Classification: The augmented feature set (original + synthetic) is used to train a classifier (a fully connected block with ReLU and dropout) while the feature extractor remains frozen.

The authors evaluated these methods on two public datasets: OpenSARShip (6 classes) and FuSARShip (9 classes), testing various class configurations (2, 4, 6, and 9 classes) to simulate different imbalance scenarios.

Key Contributions

Novel Algorithms: The paper introduces two feature-space oversampling frameworks (M2mf and M2mu) specifically adapted from the Major-to-minor (M2m) concept to address SAR ship classification imbalance.
Feature-Space Focus: Unlike traditional methods that augment raw images, this work leverages the feature space to generate synthetic data, mitigating issues related to SAR noise and resolution.
Comprehensive Evaluation: The study rigorously compares the proposed methods against a baseline (frozen extractor + classifier) and the original sample-space M2m method (M2morig) across multiple state-of-the-art backbones (ViT, VGG16, ResNet50) and dataset configurations.

Results
The experiments demonstrated that feature-space oversampling significantly outperforms both the baseline and the sample-space M2m method:

Overall Performance: The proposed methods achieved an average F1-score increase of 8.82% on the FuSARShip dataset and 4.44% on the OpenSARShip dataset compared to the baseline.
Method Comparison:
- On FuSARShip, M2mu generally yielded the best results, outperforming M2mf by 0.63% on average.
- On OpenSARShip, M2mf performed slightly better on average, outperforming M2mu by 0.98%.
- The original sample-space M2morig consistently underperformed, often yielding lower scores than the baseline, particularly in datasets with 4 or more classes.
Model Sensitivity: ViT and VGG16 showed improved performance with oversampling. In contrast, ResNet50 frequently exhibited signs of underfitting across all scenarios, suggesting its deeper architecture may be less suitable for the limited SAR datasets used in this study.
Class Configuration: The effectiveness of the methods varied by class count. For instance, M2mu excelled in 4- and 9-class FuSARShip configurations, while M2mf was superior in 2- and 6-class OpenSARShip configurations.

Significance and Claims
The paper claims that feature-space oversampling is a cost-effective and robust solution for SAR ship classification. By operating in the feature space, the approach eliminates the need to re-run the feature extraction process for the entire architecture during augmentation; only the classifier requires retraining on the synthetic features. This reduces computational resources and training time.

The authors emphasize that their work highlights the limitations of sample-space oversampling for SAR data and validates the efficacy of feature-level interventions. They conclude that while their methods significantly improve performance on imbalanced SAR datasets, future work should focus on developing SAR-dedicated sample-space oversampling techniques and further refining feature-space methods for broader application.