Class Incremental Learning With Few-Shots Based on Linear Programming for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Earth Observation
• Machine Learning

Background:

• Hyperspectral imaging (HSI) classification faces challenges in the big data era due to continuous emergence of new data classes and refined categories.
• Timeliness of application is often limited by the need for extensive retraining with new data.

Purpose of the Study:

• To develop an HSI classification model capable of rapidly learning new classifying capabilities with few examples (few-shot learning).
• To maintain high performance on original classes while adapting to new datasets.
• To enable fine-grained classification using pre-trained coarse-grained models.

Main Methods:

• Proposed a linear programming incremental learning classifier (LPILC) for adapting existing deep learning models to new datasets.
• LPILC learns new class abilities using a single example (one-shot) without requiring original class data.
• Implemented fine-grained classification by integrating LPILC with a pre-trained coarse-grained classification model.

Main Results:

• LPILC demonstrated effective adaptation to new hyperspectral datasets with minimal data, computational resources, and time.
• Achieved superior performance compared to state-of-the-art methods under identical data access and computational constraints.
• Successfully applied LPILC for fine-grained classification tasks.

Conclusions:

• LPILC offers an efficient and effective solution for incremental learning in HSI classification, suitable for time-sensitive applications.
• The proposed method can be integrated with various deep learning architectures, advancing incremental learning in remote sensing.