🤖 AI

Optical-Guided Neural Collapse for SAR Few-Shot Class Incremental Learning

This paper proposes an optical-guided framework for Synthetic Aperture Radar (SAR) Few-Shot Class Incremental Learning that leverages neural collapse and geometric priors from data-rich optical imagery to mitigate catastrophic forgetting and severe intra-class variability, achieving superior performance and feature compactness on a 24-class benchmark.

Original authors: Fan Zhang, Sijin Zheng, Fei Ma, Qiang Yin, Yongsheng Zhou, Fei Gao, Xian Sun

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fan Zhang, Sijin Zheng, Fei Ma, Qiang Yin, Yongsheng Zhou, Fei Gao, Xian Sun

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 "Forgetful Radar"

Imagine you are teaching a robot to recognize different types of airplanes using SAR (Synthetic Aperture Radar) images. SAR is like a special camera that sees through clouds and darkness, but it has a weird quirk: the same airplane looks completely different depending on the angle you look at it from. A plane seen from the front looks like a dot; from the side, it looks like a long line.

Now, imagine you want to teach this robot new planes one by one (e.g., first a fighter jet, then a cargo plane, then a helicopter). This is called Few-Shot Class-Incremental Learning. The catch? You only have a few photos of each new plane.

The Disaster: Every time the robot learns a new plane, it tends to "forget" the old ones. It's like trying to write a new chapter in a book, but the ink from the previous chapters smudges and disappears. This is called Catastrophic Forgetting.

The Clever Solution: Borrowing a Map from a Better World

The authors realized that while SAR data is scarce and confusing, Optical data (normal, high-resolution photos like you see on Google Earth) is abundant and clear. In optical photos, a plane always looks like a plane, regardless of the angle.

Their idea is simple but powerful: "Don't reinvent the wheel; borrow the blueprint."

They decided to use the "perfect" learning structure found in the optical photos to guide the messy learning process in the radar photos.

The Core Concept: "Neural Collapse" (The Perfect Classroom)

To understand their method, imagine a classroom where students (the data) are learning to sit in specific seats (classes).

The Goal: In a perfect classroom, all students from the "Math" group sit in a tight circle right next to each other, and the "History" group sits in a tight circle far away from them. The circles are perfectly spaced out, like points on a star.
Neural Collapse: This is the scientific name for when a computer model achieves this perfect arrangement. The features of the same class collapse into one tight point, and different classes spread out perfectly evenly.

The problem with SAR is that it's hard to get the students to sit in these perfect circles because there are so few of them and they look so different.

How They Did It: The "Optical Guide"

The authors built a three-step system to force the radar data to behave like the optical data:

Build the "Ideal Map" (Optical Subspaces):
First, they trained a model on thousands of clear optical photos. Because there was so much data, the model naturally figured out the perfect seating arrangement (the "Neural Collapse" geometry). They took this perfect arrangement and turned it into a set of orthogonal subspaces.
- Analogy: Think of this as drawing a perfect, rigid grid on a piece of paper using the optical photos. Each square on the grid represents a specific type of plane.
Project the Radar Data onto the Map:
When they start teaching the radar data, they don't let the radar features wander aimlessly. Instead, they force the radar features to "project" themselves onto that pre-drawn grid.
- Analogy: Imagine the radar data is a messy pile of clay. The optical grid is a cookie cutter. They press the clay into the cutter. Even if the clay is messy, it is forced to take the shape of the perfect cookie.
The Three-Part Lock (The Loss Functions):
To make sure the radar data stays in the right spot, they used three "locks" (mathematical rules) during training:
- Lock 1 (The Anchor): Keep the radar features close to the center of their assigned "cookie cutter" shape.
- Lock 2 (The Angle): Make sure the radar features stay at the correct angle relative to the other shapes (so they don't overlap).
- Lock 3 (The Center): Make sure the radar features don't wiggle around too much inside their assigned spot.

The Results: A Stable, Forget-Proof Robot

They tested this on a benchmark with 24 different target classes (mostly aircraft and military vehicles). They taught the model in sessions, adding new classes one by one.

The Competition: Other methods tried to learn the new classes but ended up forgetting the old ones, or they got confused because the radar images were too weird.
The Winner: The authors' method, which used the optical "cookie cutter" to guide the radar learning, achieved the highest accuracy.
- It didn't just learn the new planes; it remembered the old ones much better.
- The "Neural Collapse" metrics showed that the radar features actually did end up sitting in those perfect, tight circles, just like the optical ones.

Summary

In short, the paper says: "Radar data is too messy and scarce to learn perfectly on its own. So, we use clear, abundant optical photos to build a perfect geometric 'skeleton.' We then force the messy radar data to fit into that skeleton. This prevents the robot from forgetting old lessons and helps it learn new ones quickly, even with very few examples."

Technical Summary: Optical-Guided Neural Collapse for SAR Few-Shot Class Incremental Learning

Problem Statement
The paper addresses the challenge of Few-Shot Class-Incremental Learning (FSCIL) in Synthetic Aperture Radar (SAR) imagery. SAR target recognition faces unique difficulties compared to optical imagery, primarily due to severe data scarcity and high intra-class variability caused by azimuth sensitivity (the same target appearing drastically different at different angles). In an incremental learning setting, where a model must learn new classes sequentially from limited samples, these factors exacerbate catastrophic forgetting, where performance on previously learned classes degrades significantly. Existing FSCIL strategies, often designed for data-rich scenarios, struggle to maintain stable feature representations and clear decision boundaries when updating models with sparse SAR data.

Methodology
The authors propose an Optical-Guided Neural Collapse framework. The core hypothesis is that while SAR data is scarce and variable, data-rich optical imagery can provide stable geometric priors. The method leverages the theory of Neural Collapse, where optimal classification features converge to a Simplex Equiangular Tight Frame (ETF) structure (maximizing inter-class separation and minimizing intra-class variance).

The proposed framework operates in three main stages:

Orthogonal Optical Subspace Construction (Source Domain):
- A model is first trained on a large-scale, data-rich optical ATR dataset.
- Once trained, the optical feature clusters are refined into structured orthogonal subspaces. This is achieved using Singular Value Decomposition (SVD) on class-specific feature matrices to extract dominant basis vectors.
- Pairwise Alignment Scores (PAS) are computed to quantify the angular separation between these subspaces. Basis vectors that contribute most to inter-class orthogonality are weighted and retained, creating a set of orthogonal subspaces that serve as geometric proxies.
Optical-to-SAR Relation Transfer (Target Domain):
- The SAR FSCIL task utilizes a Unconstrained Features Model (UFM) architecture with a fixed Simplex ETF classifier. The classifier weights are frozen to provide stable geometric anchors.
- Base Session: SAR base classes are learned using a Dot Regression (DR) loss to align features with the fixed ETF prototypes.
- Incremental Sessions: For new SAR classes, the method employs a cross-domain alignment strategy:
  - Proxy Assignment: Each new SAR class is randomly assigned to an optical subspace (acting as a geometric proxy rather than a semantic match).
  - Subspace PA Alignment Loss ( $\mathcal{L}_{orth}$ ): This loss minimizes the Principal Angle (PA) between the SAR feature and the assigned optical orthogonal subspace, forcing SAR features to collapse into these structured partitions.
  - Subspace Center Alignment Loss ( $\mathcal{L}_{center}$ ): A regularized MSE loss aligns SAR features with the centroid of the assigned optical subspace to reduce intra-class fluctuation.
- The total loss combines the DR loss, PA alignment loss, and center alignment loss.
Implementation Details:
- The backbone is a pretrained SAM (Segment Anything Model) ViT-B encoder.
- To handle domain shift efficiently, the image encoder is frozen during SAR adaptation. Only lightweight MLP and Space Adapters, along with LoRA (Low-Rank Adaptation) modules in the projection layers, are fine-tuned.

Key Contributions

Orthogonal Subspace Construction: The paper introduces a method to construct structured, orthogonal subspaces from a well-trained optical classification model. These subspaces serve as geometric priors to guide SAR feature learning, addressing the lack of data in the SAR domain.
Optical-to-SAR Relation Transfer: A novel transfer module is proposed that aligns SAR features with the constructed optical subspaces. This is achieved through a combination of Principal Angle constraints and center alignment, effectively transferring the "neural collapse" geometry from the optical domain to the SAR domain without requiring semantic class matching.
Experimental Validation: The method is validated on a benchmark comprising an optical ATR dataset and a SAR ATR dataset (24 target classes) organized into one base session and seven incremental sessions.

Results

Performance: The proposed method achieves the highest final accuracy (76.46%) compared to state-of-the-art FSCIL methods (NC-FSCIL, PriViLege, WaRP, FACT) on the SAR benchmark.
Stability: It demonstrates a favorable trade-off between final performance and performance degradation (Performance Degradation of 23.23%, outperforming most baselines).
Neural Collapse Metrics: Quantitative analysis using neural collapse metrics ( $\Sigma$ , $std_{ETF}$ , $mean_{ETF}$ , $SD$) shows that the proposed method induces better intra-class compactness and inter-class separability. The features more closely approximate the ideal simplex-ETF geometry compared to baselines, particularly in later incremental sessions.
Ablation: Removing the orthogonal alignment losses leads to significant performance drops and higher feature variance, confirming the necessity of the optical-guided geometric constraints.

Significance
The paper claims that its primary significance lies in providing a practical mechanism to transfer structured geometric knowledge from data-rich domains (optical) to data-scarce domains (SAR). By using orthogonal subspaces as intermediate anchors, the method bridges the gap between the idealized neural collapse theory and the practical constraints of few-shot incremental learning. This approach stabilizes the learning process, mitigates catastrophic forgetting, and enables the model to learn robust representations for SAR targets despite severe azimuth variability and limited training samples. The work suggests that enforcing orthogonality is a more achievable and structured path to neural collapse in real-world, imbalanced scenarios than strictly converging to a simplex ETF directly.