Dr. Yuxiang Zhang and Prof. Shunlin Liang’s recent publication in the ISPRS Journal of Photogrammetry and Remote Sensing

SpectralX: Parameter-efficient domain generalization for spectral Remote Sensing Foundation Models

Congratulations to Yuxiang Zhang, Prof. Shunlin Liang, and the research team on publishing a new article in the ISPRS Journal of Photogrammetry and Remote Sensing (Vol. 239, pp. 774-792). The article is titled “SpectralX: Parameter-efficient domain generalization for spectral Remote Sensing Foundation Models” and is available at https://doi.org/10.1016/j.isprsjprs.2026.06.008.

In recent years, research on Remote Sensing Foundation Models (RSFMs) has achieved remarkable progress. A variety of models have been pre-trained on large volumes of optical image data and have demonstrated outstanding performance across diverse downstream tasks. However, due to the difficulty of acquiring high-quality spectral data and the associated large data volumes, there remains a notable absence of foundation models specifically designed for multispectral/hyperspectral data. In this paper, we propose SpectralX, a lightweight fine-tuning method that adapts optical RSFMs to the spectral modality with only a small number of trainable parameters, significantly enhancing domain generalization performance using limited annotated data.

The main framework of SpectralX consists of three stages: Spectral Modality Adaptation (Stage 1), Task-Oriented Generalization Training (Stage 2), and Unseen Scene Interpretation (Stage 3). In Stage 1, the training objective is a masked reconstruction task. We adopt foundation models pre-trained on optical data (Scale-MAE and SatMAE++) as the backbone, freeze all backbone parameters, and update only a small number of parameters in HyperT and AoMoA. Among these, AoMoA is inserted exclusively into four specific layers of both the encoder and decoder, enabling efficient parameter fine-tuning and cross-modal feature adaptation. In Stage 2, the training objective is a downstream task (segmentation). We inherit the encoder architecture from Stage 1 (with the decoder removed), add an Area-adapter after AoMoA, and employ UperNet as the segmentation head to guide the model toward adapting to the downstream task and generalizing to unseen scenes. In Stage 3, the SpectralX fine-tuned model is directly used to interpret observation scenes that were not encountered during training.

The three-stage framework of SpectralX, featuring Spectral Modality Adaptation, Task-Oriented Generalization Training, and Unseen Scene Interpretation

Reference:

Yuxiang Zhang, Wei Li, Mengmeng Zhang, Jiawei Han, Ran Tao, Shunlin Liang, SpectralX: Parameter-efficient domain generalization for spectral Remote Sensing Foundation Models, ISPRS Journal of Photogrammetry and Remote Sensing, Volume 239, 2026, Pages 774-792, ISSN 0924-2716, https://doi.org/10.1016/j.isprsjprs.2026.06.008.

Keywords:  Multispectral image; Hyperspectral image; Domain generalization; Parameter-efficient fine-tuning; Foundation model