all dataset

Citation Author(s):
Xiaomin
Tian
Submitted by:
Xiaomin Tian
Last updated:
Tue, 04/01/2025 - 23:10
DOI:
10.21227/rhv1-9b44
License:
0
0 ratings - Please login to submit your rating.

Abstract 

Hyper-spectral unmixing is a technique to estimate the abundances of different endmembers in each mixed pixel of remote sensing images. Deep learning has made significant progress in this area, offering automatic feature extraction, nonlinear pattern recognition, and end-to-end solutions. However, existing deep learning models have not fully utilized the spectral information of endmembers, leading to insufficient data mining. We propose a pixel-based Additive Attention Neural Network (AANN) that uses endmember spectral feature vectors as auxiliary data for training, which helps improve the accuracy of mixed pixel decomposition. Additionally, to validate the influence of adjacent pixels, we developed a spatial-based AANN that adds a convolutional layer to extract spatial features, exploring the endmember decomposition accuracy for various window sizes. Experimental results show: 1. The pixel-based AANN significantly outperformed traditional machine learning and model-based methods; 2. The spatial-based AANN, which incorporated features of neighboring pixels, performed nearly as well as the pixel-based AANN on the Jasper Ridge dataset, but showed a notable 40.5% improvement on the urban dataset; 3. The spatial-based AANN showed the best effect with a 3 × 3 window size. These results indicate that adding spectral information and adjacent pixel information can effectively improve unmixing accuracy.

Instructions: 

Jasper Ridge is a popular hyperspectral data used in [enviTutorials, SS-NMF, DgS-NMF,RRLbS, L1-CENMF]. There are 512 x 614 pixels in it. Each pixel is recorded at 224 channels ranging from 380 nm to 2500 nm. The spectral resolution is up to 9.46nm. Since this hyperspectral image is too complex to get the ground truth, we consider a subimage of 100 x 100  pixels. The first pixel starts from the (105,269)-th pixel in the original image. After removing the channels 1--3, 108--112, 154--166 and 220--224 (due to dense water vapor and atmospheric effects), we remain 198 channels (this is a common preprocess for HU analyses). There are four endmembers latent in this data: "#1 Road", "#2 Soil", "#3 Water" and "#4 Tree".

Urban is one of the most widely used hyperspectral data used in the hyperspectral unmixing study. There are 307 x 307 pixels, each of which corresponds to a 2 x 2 m2area. In this image, there are 210 wavelengths ranging from 400 nm  to 2500 nm, resulting in a spectral resolution of 10 nm. After the channels 1--4, 76, 87, 101--111, 136--153 and 198--210 are removed (due to dense water vapor and atmospheric effects), we remain 162  channels (this is a common preprocess for hyperspectral unmixing analyses). There are three versions of ground truth, which contain 4, 5 and 6 endmembers respectively, which are introduced in the ground truth.