Comparison of Satellite Hyperspectral Data in Geological Mapping (Methods Part 1)

exist Comparison of satellite hyperspectral data in geological mapping (Data section)The previous sections introduced the study area, data from different sensors, and data preprocessing. This section mainly introduces the relevant research techniques.

1. Data Registration

Data registration is the process of precisely aligning datasets from different sources to enable direct comparison and analysis within the same spatial reference frame. In this study, to compare satellite multispectral data from HyMap, EnMAP, PRISMA, and EMIT sensors, the following steps were performed for data registration:

1) Data subset selection: First, for the data from each sensor, Rupsa et al. selected a subset region corresponding to the study area.

2) Using AROSICS tools: Utilize the open-source image registration software (AROSICS, version 1.11.0) to automatically identify the correlation between comparable bands and calculate key points for data registration.

2. Global Affine Transformation

A global affine transformation is applied to adjust and resample the data so that all datasets are reprojected onto the same grid as the HyMap data. This step is accomplished using the Rasterio and scikit-image libraries.

3. Resampling

To minimize the impact of resampling on the final results, Rupsa et al. performed data resampling as late as possible in the last stage of the analysis process.

The above steps ensure spatial consistency between different sensor datasets, which is crucial for accurate qualitative and quantitative analysis. Data registration allows for the comparison of spectral characteristics across different datasets, assessment of their application potential in geological mapping, and identification of consistency and differences between different sensors.

4. Spectral analysis

Spectral analysis is an important technique in the field of remote sensing for identifying and classifying surface materials. In this study, researchers used two main spectral analysis methods to compare and evaluate the application of different satellite sensor data in geological mapping.

4.1 Spectral Index

Spectral index analysis generates an image by dividing one band by another. This image is relatively insensitive to many common noise effects, such as atmospheric effects, because these effects are usually largely canceled out during the division process. Typically, the bands are chosen such that the denominator represents the minimum value of the target absorption feature, and the numerator is the reflectance associated with one or two adjacent bands. This helps distinguish mineral types because it cancels out effects between largely identical bands while enhancing local variations caused by mineral absorption.

Rupsa et al. used several established spectral indices to analyze minerals in two study areas. In Eppembe, the spectral indices were used to distinguish calcium carbonate-rich rock gullies from the surrounding altered host rock. For Marinkas-Quellen, the spectral indices were used to map the distribution of calcium carbonate and magnesium carbonate rocks. These formulas and their sources are described in Table 1.

The main steps of spectral index analysis:

(1) Select a specific band that is relevant to the target mineral.

(2) Calculate the spectral index value to enhance the detection of target minerals.

(3) Use these index values ​​to generate geological feature maps.

Table 1. Formulas for spectral indices used in the study.

4.2 Spectral Abundance

Spectral unmixing refers to the process where the signal from a spectral pixel is a linearly weighted combination of the abundance of endmember spectra. Endmembers are certain pristine pixels in an image that represent the spectral characteristics of approximately pure substances. Through linear spectral unmixing, the abundance of various substances in an image can be estimated. These abundances can be represented as spectral abundance maps, reflecting the latent variables of spectral variation (i.e., hidden variables that determine the variation of spectral pixels). Latent variables are quantities that cannot be directly observed; they represent hidden aspects of the data and are inferred through mathematical modeling and relationships between observable variables. This approach has limitations in some cases: pristine pixels in satellite data may be uncommon due to coarse spatial resolution, and some substances may be mixed in a scale-dependent nonlinear manner. However, representing hyperspectral images as sparse combinations of weighted pristine pixel endmembers provides a method for comparing latent information. Therefore, by manually defining fixed endmember pixel locations for each study area and assuming these locations represent typical “pure spectra” for each geological type in the scene, a corresponding endmember spectral library was extracted for each sensor using these endmember locations.

Rupsa et al. used EnMAP as a reference and selected endmember pixels because their sampled spectra were most similar to those of other sensors. Linear spectral unmixing was performed using non-negative least squares to extract abundance maps from each image. Due to the different spectral responses, Rupsa et al. selected different spectral subsets for unmixing in the Marinkas-Quellen and Epembe regions. For Epembe, the VNIR range of 450–1150 nm and the SWIR range of 2000–2480 nm were considered because characteristic absorptions of iron precipitates and carbonates exist in these ranges. For Marinkas-Quellen, only the SWIR range of 2000–2480 nm was considered because the dominant minerals here are carbonates. Furthermore, water absorption bands were excluded during unmixing because these bands typically have high noise levels in remote sensing data.

In this way, Rupsa et al. generated abundance maps and used them to compare the consistency between different sensors. They found that even with significant spectral biases in the endmember spectra, the abundance maps from various sensors exhibited similar patterns of geological change. This suggests that despite the consistency issues between spectral data, unmixing techniques can reveal underlying features consistent with geological changes. This comparison allows for the assessment of the different sensors' ability to capture geological features and the identification of which sensors are more suitable for geological mapping.

 参考：[1]:Chakraborty, R.; Rachdi, I.; Thiele, S.; Booysen, R.; Kirsch, M.; Lorenz, S.; Gloaguen, R.; Sebari, I. A Spectral and Spatial Comparison of Satellite-Based Hyperspectral Data for Geological Mapping. Remote Sens. 2024, 16, 2089. https://doi.org/ 10.3390/rs16122089