How do hyperspectral satellites "see through" vegetation diversity? — Hybrid pixel decomposition: a technology for identifying multiple vegetation types from a single pixel.

Traditional species diversity surveys rely on ground quadrats, which are limited by manpower, transportation, and terrain conditions, making it difficult to achieve large-area coverage. Although the relevant technical specifications of the Ministry of Ecology and Environment have designed diversity assessment methods, this indicator has long been lacking in actual comprehensive assessments due to the difficulty in obtaining ground observation data. We need a technology that can "see" vegetation diversity from the sky.

The entire technical route consists of four core steps:

Step 1: Determine "how many types of vegetation are there" (virtual dimension estimation)

The HFC algorithm was used to automatically determine the number of independent vegetation signal sources in the image through statistical hypothesis testing. In this analysis, seven endmembers were automatically identified—representing seven different vegetation types or states in the Qinling Mountains region.

Step 2: Extract "what each type of vegetation looks like" (endmember extraction)

Using the VCA (Vertex Component Analysis) algorithm, seven of the purest spectral features were automatically identified from millions of pixels. These endmembers may correspond to different vegetation types such as evergreen coniferous forests, deciduous broad-leaved forests, shrublands, mixed forests, and farmland.

Step 3: Calculate "how much of each type"Abundance Inversion）

For each pixel, the abundance of each endmember is calculated using the FCLS (Fully Constrained Least Squares) algorithm. The constraint is that all abundances are non-negative and sum to 1—ensuring the physical meaning is reasonable. This results in seven abundance distribution maps, each representing the spatial distribution density of a particular vegetation type.

Figure 1.Spatial distribution of abundance of 7 vegetation endmembers (each pixel is decomposed into multiple components)

Step 4: Calculate the diversity index

The abundance within each pixel is considered as "species composition," and the classic diversity index, the Shannon-Wiener index (which combines richness and evenness), is calculated.Simpson index(The probability of randomly selecting two individuals belonging to different types), richness (number of effective endmembers), and evenness (whether the distribution is balanced). Finally, a weighted fusion is performed to obtain the comprehensive biodiversity index BDI (0-1, higher is better diversity).

Key advantages: High-resolution satellites are suitable for detecting patch changes, and multispectral satellites can acquire vegetation vitality indicators such as NDVI, but neither can resolve the vegetation type composition within a pixel at the spectral level. Hyperspectral unmixing technology fills this gap—it provides quantitative information on vegetation diversity at the sub-pixel level, which is currently unavailable from other types of remote sensing satellites.

This analysis of an area of ​​approximately 6400 km² (about 4 million effective vegetation pixels) in the hinterland of a mountainous region reveals:

1. Overall diversity is moderate to high.The mean BDI is 0.568, and most pixels are composed of a mixture of three vegetation end-members (accounting for 65.5%), reflecting the vegetation heterogeneity in the region.

2. "Green" does not equal "diversity"The correlation between biodiversity index (BDI) and vegetation vitality index (NDVI) was only weak (r=0.369), demonstrating that areas with abundant vegetation cover are not necessarily species-rich (e.g., pure coniferous forests have high NDVI but low biodiversity). This validates the unique value of the decontamination method.

Figure 2.Pearson correlation coefficient matrix of diversity indicators

Figure 3.NDVI vegetation index distribution map

Figure 4.Spatial distribution of the Composite Biodiversity Index (BDI) – Green indicates areas of high diversity.

3. The elevation gradient pattern is clear.：CombinationSRTM DEM dataDiversity decreases significantly with altitude (r=-0.623), exhibiting a three-tiered pattern.

Low altitude (Deciduous broad-leaved and evergreen mixed forest, BDI=0.65~0.75, with the highest diversity;

Mid-altitude (1000-1800m)Transitional zone: biodiversity declines rapidly, while the proportion of coniferous forests increases;

High altitude (>1800m)Pure coniferous forests dominated by fir and redwood have a BDI of 0.13 to 0.21 and low but stable biodiversity.

Figure 5.DEM distribution in the study area

Figure 6.Trends in diversity indicators with altitude gradient

Figure 7.The negative correlation between BDI and altitude (r=-0.623)

Hyperspectral mixed pixel decomposition has opened up a completely new path for monitoring vegetation diversity over large areas. Compared with traditional quadrat surveys, it has a wide coverage (a single scene can cover thousands of square kilometers) and high timeliness (analysis can be completed in a few hours); compared with high-resolution and multispectral satellites, it has the unique ability to distinguish vegetation composition from the spectral dimension, and can directly produce quantitative indicators of diversity, rather than just providing indirect information on "greenness".

In the future, combining multi-temporal data will enable the monitoring of diverse interannual dynamic changes, and ground-based quadrat validation will allow for the establishment of [a system/mechanism].Spectral diversityThe quantitative relationship between biodiversity and species diversity. This method is expected to fill the long-standing gap in the national ecological function assessment of the lack of "diversity" indicators, and provide objective and repeatable technical support for the quantitative assessment of the effectiveness of ecological governance.