HyperVision: A Channel-Adaptive Ground-Based Hyperspectral Vision Pre-trained Backbone

This is the official repository for the paper "HyperVision: A Channel-Adaptive Ground-Based Hyperspectral Vision Pre-trained Backbone". https://arxiv.org/abs/2605.17286

and Hyperspectral Images Reader with multiple dataset support.

Looking for the dataset readers? The new HSI Reader project is a collection of readers for ground-based hyperspectral datasets. It retains V1 support and adds 28 datasets; we recommend starting there for standalone dataset reading.

Code: https://github.com/lronkitty/HyperVision

📢 Announcement: Our paper "HyperVision: A Channel-Adaptive Ground-Based Hyperspectral Vision Pre-trained Backbone" has been officially accepted by BMVC 2026! 🎉

Official PyTorch implementation for the BMVC 2026 paper: "HyperVision: A Channel-Adaptive Ground-Based Hyperspectral Vision Pre-trained Backbone"

---

🚀 News

• [31-Jul-2026] 🎉 HyperVision has been officially accepted to BMVC 2026!
• [02-Jul-2026] 📦 Official codebase, model checkpoints, and inference pipelines released.

---

📖 Abstract

While hyperspectral imaging provides rich spatial-spectral information across hundreds of narrow wavelength bands for precise material identification, ground-based hyperspectral pre-trained backbones remain absent, constrained by varying spectral configurations across sensors, limited annotations and heterogeneous labeling schemes, and the limited scale and scene diversity of existing datasets. To address these challenges and enable universal perception, we propose HyperVision, the first ground-based hyperspectral pre-trained backbone. First, to handle varying spectral configurations, HyperVision adopts a channel-adaptive dynamic embedding mechanism to map heterogeneous inputs into a unified token space. Second, we develop an unsupervised representation learning framework. Specifically, to address limited annotations and heterogeneous labeling schemes, a multi-source pseudo-labeling method is introduced to fuse spatial structures from SAM2 and fine-grained spectral material information from HyperFree. Furthermore, to enrich scene diversity and compensate for limited dataset scale, a cross-modal knowledge distillation mechanism is utilized to transfer rich semantic representations from a pre-trained RGB vision model to our backbone. Pre-trained on a collection of 15k images from 26 diverse ground-based datasets, HyperVision demonstrates exceptional generalization. Requiring only efficient head-only adaptation without adjusting backbone parameters, it outperforms state-of-the-art task-specific methods across three downstream tasks under varying sensor configurations, yielding up to a 16.3% relative improvement in hyperspectral semantic segmentation $\mathrm{Acc}_{\mathrm{M}}$, a 2.1% relative gain in object tracking AUC, and a 35.5% reduction in salient object detection MAE.

---

🛠️ Main Code (HyperVision)

The core architecture and model initialization utilities are located under the HyperVision directory.

To initialize the model, you can use the model builder functions from HyperVision/build_HyperVision.py. We support different model configurations:

• HyperVision-B: build_HyperVision_b
• HyperVision-L: build_HyperVision_l
• HyperVision-H (Default): build_HyperVision_h

Python Example:

import torch
from HyperVision import build_HyperVision_h, build_HyperVision_l, build_HyperVision_b
from hyperspectral_image_reader import load_hypervision_matrix, LoadHyperspectralImage

device = "cuda" if torch.cuda.is_available() else "cpu"
image_size = 512

# 1. Build the HyperVision model (HyperVision-H configuration shown below)
model = build_HyperVision_h(
    checkpoint="path/to/hypervision_distilled_h.pth",             # Optional checkpoint path (.pth weights)
    image_size=512,                                         # Input resolution
    vit_patch_size=16,                                      # Patch size of the backbone
    encoder_global_attn_indexes=[15, 23, 31],               # Layers using global attention
    merge_indexs=[8, 32],                                   # Layers doing patch merging
    class_number=-1                                         # Number of classes for mask decoder
)
model = model.to(device)
model.eval()

# 2. Load and preprocess a hyperspectral image
# Preprocessed shape: (H_ori, W_ori, C_hsi) with values in range [0, 255]
img_matrix = load_hypervision_matrix("path/to/image.mat", dataset_name="harvard")
h_ori, w_ori = img_matrix.shape[:2]

# Convert to torch tensor with shape (1, C_hsi, H_ori, W_ori)
input_tensor = torch.as_tensor(img_matrix, device=device).permute(2, 0, 1).unsqueeze(0).float()

# 3. Automatically retrieve the wavelengths list from the dataset pipeline config
loader = LoadHyperspectralImage(dataset_type="harvard")
wavelengths = loader.wavelengths

# 4. Extract features via the HyperVision Image Encoder
with torch.no_grad():
    # GSD represents Ground Sampling Distance (defaults to [0.01])
    multi_stage_features = model.image_encoder(
        input_tensor, 
        test_mode=True, 
        input_wavelength=wavelengths, 
        GSD=[0.01]
    )

---

📊 Dataset Reading & Processing

We provide read_dataset_image.py to read and preprocess hyperspectral images from various datasets. The script utilizes the custom data pipeline class LoadHyperspectralImage from hyperspectral_image_reader/hyperspectral_pipelines.py to normalize the images (scaled to [0, 255]) and reshape them into the format expected by HyperVision: (H_ori, W_ori, C_hsi).

Python API Usage:

from hyperspectral_image_reader import load_hypervision_matrix

# Load and preprocess HSI matrix
# dataset_name must be one of the supported dataset identifiers (e.g., 'harvard', 'arad_1k_31')
img_matrix = load_hypervision_matrix("path/to/image.mat", dataset_name="harvard")
print("Processed shape (H, W, C):", img_matrix.shape)

Command Line Interface:

You can also run the script from the command line to load an HSI image and optionally save the processed matrix as a .npy file:

python hyperspectral_image_reader/read_dataset_image.py --path /path/to/image.mat --dataset harvard --output output.npy

---

🗃️ Supported Datasets

Below is a summary of the 26 ground-based hyperspectral datasets (31 subset entries) supported by our pipeline, along with their respective keys for the --dataset parameter:

Wavelength and decoding notes: HSSOD contains 81 bands over 380–780 nm at 5 nm spacing. The HOT-2024 RedNIR loader decodes its 4×4 PNG mosaic and returns 15 channels; its listed 600–850 nm centers are nominal estimates. The output channel count alone does not establish whether a channel is all zero.

---

📝 Citation

If you find our work or this project helpful, please consider citing our paper:

@misc{fu2026hypervisionarxiv,
  title         = {{HyperVision: A Channel-Adaptive Ground-Based Hyperspectral Vision Pre-trained Backbone}}, 
  author        = {Guanyiman Fu and Jingtao Li and Zihang Cheng and Zhuanfeng Li and Diqi Chen and Yan Xu and Xiangyu Liu and Fengchao Xiong and Jianfeng Lu and Chengrong Chen and Jun Zhou},
  year          = {2026},
  eprint        = {2605.17286},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.17286}
}