Title: MixerSENet: A Lightweight Framework for Efficient Hyperspectral Image Classification

Abstract:

This study presents MixerSENet, a novel architecture designed to tackle the dual challenges of high computational demands and the scarcity of labeled data in hyperspectral image (HSI) classification. The framework operates on HSI patches, ensuring that spatial and channel dimensions are effectively decoupled while preserving uniform size and resolution across all network layers. As a lightweight solution, MixerSENet is highly computationally efficient and requires significantly fewer parameters than conventional models, rendering it ideal for deployment in resource-constrained settings. To further boost feature extraction capabilities, the model integrates a squeeze and excitation block, which helps the network identify and leverage more informative features.

Benchmark evaluations on two datasets highlight MixerSENet’s superior performance, achieving an overall accuracy (OA) of 82.47% on the Houston13 dataset and 96.70% on the Qingyun dataset. These results surpass those of leading state-of-the-art methods, such as 3D-CNN, HybridKAN, HSIFormer, SimPoolFormer, and MorphMamba. Detailed analysis of computational efficiency reveals that MixerSENet strikes an optimal balance between precision and speed, utilizing only 53,146 parameters and demonstrating low inference time. This combination confirms the model’s practicality for real-world applications. The source code for MixerSENet will be made publicly available upon publication at https://github.com/mqalkhatib/MixerSENet.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC