Multi-Teacher Multi-Objective Meta-Learning for Zero-Shot Hyperspectral Band Selection

TL;DR

This work tackles zero-shot hyperspectral band selection by addressing the lack of cross-dataset generalization in existing methods. It introduces M$^3$BS, a framework that combines a spatial-spectral graph with a generalizable GCN parameterized by dataset-agnostic bases and dataset-specific coefficients, augmented by multiple band-selection teachers and an uncertainty-based multi-objective meta-learning objective for joint band selection and classification. The key contributions include the dataset-agnostic meta-knowledge extractor, a diversity ensemble of teachers to enhance generalization, and an end-to-end training procedure that transfers to unseen datasets without retraining. Experiments on six hyperspectral datasets show that M$^3$BS achieves strong zero-shot performance and outperforms several baselines, demonstrating both generalization and efficiency benefits for practical hyperspectral analysis.

Abstract

Band selection plays a crucial role in hyperspectral image classification by removing redundant and noisy bands and retaining discriminative ones. However, most existing deep learning-based methods are aimed at dealing with a specific band selection dataset, and need to retrain parameters for new datasets, which significantly limits their generalizability.To address this issue, a novel multi-teacher multi-objective meta-learning network (M$^3$BS) is proposed for zero-shot hyperspectral band selection. In M$^3$BS, a generalizable graph convolution network (GCN) is constructed to generate dataset-agnostic base, and extract compatible meta-knowledge from multiple band selection tasks. To enhance the ability of meta-knowledge extraction, multiple band selection teachers are introduced to provide diverse high-quality experiences.strategy Finally, subsequent classification tasks are attached and jointly optimized with multi-teacher band selection tasks through multi-objective meta-learning in an end-to-end trainable way. Multi-objective meta-learning guarantees to coordinate diverse optimization objectives automatically and adapt to various datasets simultaneously. Once the optimization is accomplished, the acquired meta-knowledge can be directly transferred to unseen datasets without any retraining or fine-tuning. Experimental results demonstrate the effectiveness and efficiency of our proposed method on par with state-of-the-art baselines for zero-shot hyperspectral band selection.

Figures (4)

• Figure 1: The overall architecture of the proposed M$^3$BS for zero-shot hyperspectral band selection.
• Figure 2: The first graph convolutional layer inside the generalizable GCN, comprising consecutive matrix multiplications. The weight matrix is parameterized as dataset-agnostic bases and dataset-specific coefficients. These bases are shared by different datasets, while these coefficients, which are specific to each dataset, are computed by a simple MLP network.
• Figure 3: The multi-objective multi-objective meta-learning procedure for zero-shot optimization, where different colors correspond to different HSI datasets.
• Figure 4: The false-color images composited by three different spectral bands from (a) Indian Pines, (b) Pavia University, and (c) University of Houston.