This repository provides a PyTorch implementation of the SpecAware model. Paper Link

Pretrained weights are available here.

For each input hyperspectral image, the corresponding sensor and dataset parameters are required, including wavelength, FWHM, GSD, sensor name, and processing level.

wavelength = [365.9298, 375.5940, 385.2625, ...2496.236]
FWHM = [9.85211, 9.79698, 9.74410, ...9.99943]
GSD = 10.0
name = "others"
level = "L1"

A Python environment used in our experiments:

python    = 3.11.11
pytorch   = 2.5.1
torchdata = 0.11.0
timm      = 1.0.15
sentence-transformers = 4.1.0
gdal      = 3.6.2

All data used to construct the pre-training dataset were obtained from the official AVIRIS data portal. We provide a file manifest and some scripts under Hyper-400K/.

You can download the raw AVIRIS products from the official portal by following the entries in the manifest.

An example subset is available here.

Please cite our paper if our work is helpful for your research.

[1] Ji, R., Wang, X., Niu, C., Zhang, W., Mei, Y., Tan, K., 2026. SpecAware: a spectral-content aware foundation model for unifying multi-sensor learning in hyperspectral remote sensing mapping. ISPRS J. Photogramm. Remote Sens. 234, 242–260.

For questions or feedback, please contact: ecnu.jirenjie@gmail.com.

This research benefited from the following resources: MAE, DOFA, Copernicus-FM, Scale-MAE, Aurora, and SMP.