This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).
Hyperspectral images represent an important source of information to assess ecosystem biodiversity. In particular, plant species richness is a primary indicator of biodiversity. This paper uses spectral variance to predict vegetation richness, known as Spectral Variation Hypothesis. Hierarchical agglomerative clustering is our primary tool to retrieve clusters whose Shannon entropy should reflect species richness on a given zone. However, in a high spectral mixing scenario, an additional unmixing step, just before entropy computation, is required; cluster centroids are enough for the unmixing process. Entropies computed using the proposed method correlate well with the ones calculated directly from synthetic and field data.
Hyperspectral images represent an important tool to assess ecosystem biodiversity. Improvements in spectral and spatial sensors resolution allow more precise analysis of biodiversity indicators that should agree with indicators obtained using field data. The development of accurate analysis tools would be advantageous to extend the analysis to larger zones by hyperspectral image processing; furthermore, for future scenarios, given that the actual data gathering is expensive in time and resources, a preanalysis would give extra knowledge for planning a more effective campaign to collect field data.
One of the most important indicators of biodiversity in ecosystems is plant species richness. This indicator can be measured in hyperspectral images considering the Spectral Variation Hypothesis (SVH) proposed in [
The main contribution of this paper is the focus on proportional abundances that entropy indexes can capture which are more meaningful than species richness alone. It also evaluates the use of hierarchical clustering to assess biodiversity and the use of spectral unmixing to improve results.
This paper is organized as follows: Section 2 gives a brief theoretical background. Section 3 explains the methodology used to process hyperspectral imagery. Section 4 explains the details of experiments and gives the results. Section 5 discusses the experimental results. Section 6 presents the conclusions.
The goal of Hierarchical Clustering is to build a hierarchy of clusters, a cluster tree. There are two main approaches to build a hierarchy of clusters: Agglomerative and divisive. The former approach is used in this work. Hierarchical Agglomerative Clustering (HAC) is a “bottom up” approach in which, at first, each observation constitutes its own cluster; then, pairs of clusters are merged successively, generating the hierarchy, until there is just one cluster left. HAC needs two concepts to carry out the process: a metric (distance) and a linkage [
Metrics used to construct the dissimilarity matrix:
Euclidean distance:
Spectral angle distance (SAD):
The linkage used in this work is known as completelinkage [
Every time HAC has to merge two clusters, it chooses clusters that have minimal distance. Merging n observations requires n−1 operations. This n−1 linkage valued sequence produce a characteristic “elbow” curve. Obtaining a number of clusters in a nonsupervised fashion requires choosing a point in the “elbow” part of the curve. The Number of Clusters (NC) is determined using the method in [
The Shannon entropy is a popular biodiversity index that represents the weighted geometric mean of the proportional abundances of the species [
Pearson's correlation coefficient between two variables is defined as the covariance of the two variables divided by the product of their standard deviations [
In a hyperspectral image, the reflected or emitted radiation measured per pixel, rarely comes from the interaction with a single homogenous material. If the spatial resolution is higher than the size of an object in the image, it is very likely that the image have pure pixels, though this case is not frequent.
There are two models to analyze the mixing problem, linear model and nonlinear model. For simplicity, only the linear mixing model [
Frequently, there are spectral signature differences, even among pure pixels of the same material, attributable to shadows or different reflection angles that weaken measurements; the linear mixing model accounts these cases with abundances in
The Spectral Unmixing (SU) approach used in this study computes matrix
Since the proposed method is aimed at capturing the Shannon entropy for a set of pixels. The ideal case occurs when all pixels are pure ones; in this case, each cluster represents specific species clearly. The typical case is when a pixel represents not one, but a combination of different species; in this case, clusters are less useful to capture species diversity since several species can be grouped in the same cluster. The unmixing process decomposes the spectral combination of species into abundances for every pixel; these abundances can be used as in the ideal case to compute entropy.
In the following sections, terms like high, mid or low spectral mixing are used. These terms refer to the proportion in spectral data between pixels that are a combination of endmembers
This stage prepares the hyperspectral image by removing duplicate pixels, and nonvegetation pixels. Duplicate pixels can result from the nearestneighbor resampling of image pixels during some geo correcting procedures. Duplicates removal is efficiently achieved using a radixsort based algorithm [
The hyperspectral image is partitioned in a set of rectangular parts. It is not mandatory that all rectangles have the same dimensions. However, it is advisable that the number of pixels contained in rectangular elements is as uniform as possible; this ensures that possible entropy values are in the same range, making comparisons between entropy values, in different rectangular elements, more meaningful.
This algorithm constitutes the core procedure of the proposed method. The algorithm deals successfully with a general spectral mixing scenario, under the requirement that the number of endmembers is provided.
Input:
Output:
For each partition element
Step 1. Apply HAC to
Step 2. Apply NC to
Step 3. From
Step 4. From
Step 5. Compute Shannon entropy
There is a variant to this algorithm that could be used in particular cases, when data exhibits low or mid spectral mixing. The benefit of the variant is that spectral unmixing is completely avoided; resorting only to cluster information to compute entropy and doing so in a truly automatic fashion. Unfortunately, deciding when to use this variant is very circumstantial requiring expert human judgment to apply it. Following is the variant, which reuses most of the notation of the previous algorithm.
Input:
Output:
For each partition element
Step 1. Apply HAC to
Step 2. Apply NC to
Step 3. From
Step 4. Compute Shannon entropy
Every part in the original partition is replaced with its respective Shannon entropy index; the value is replicated to match the original size of the partition element, then the new image is reassembled as shown in
The experiments tried to verify that the proposed method computes a biodiversity index, which reflects to a certain degree the species richness in the hyperspectral image. Two kinds of experiments are conducted; the first experiment uses synthetic hyperspectral data to assess the influence of spectral mixing, to verify proper operation and to identify limitations of the proposed method. The spectral mixing degree refers to the portion of nonpure pixels in the dataset. The second experiment uses real hyperspectral data along with field data to assess the index of correlation between the two datasets.
Data is generated assuming a linear mixing model. Endmembers spectra come from an independent unmixing process. All endmembers represent vegetation. Considering the notation introduced in Section 2.7,
Real entropy index of synthetic pixels is computed using matrix
The data is from the Guanica dry forest in Puerto Rico, the field data is from 25 circular plots, each plot has a radius of 10 m. The sampling was conducted between 2006 and 2009; sample locations were chosen at random and stratified by history of land use and topographic position. Field data considered vegetation, specifically plants with stem radius larger than 5 cm. The corresponding hyperspectral images are from AISA airborne imagery acquired in 2007, the images have 128 bands, wavelengths from 400–1000 nm, and spatial resolution of 1 m. The data is divided in four groups, because there are four separate hyperspectral images taken at different times producing differences in reflectance.
Under these specifications, every hyperspectral zone had 324 pixels, where every pixel is a vector of 128 components.
The number of endmembers for the unmixing stage is set up corresponding to the species richness of the plot, also collected with field data. The proposed method processes the whole data batch in 30 s approximately. The validation of the proposed index is done calculating the Pearson correlation coefficient between the Shannon entropy of actual field data and the Shannon entropy computed in the corresponding image using the algorithm in Section 3.3.
The effect of spectral mixing over HAC can be better appreciated in a simplified hypothetical example, using only three endmembers, one hundred points, two bands and two degrees of spectral mixing, as depicted in
First experiment confirmed how degree of spectral mixing of data affects the proposed method. HAC gives results that are acceptable or even good with a medium mixing degree or less. The worst case appears in a high spectral mixing scenario, where important clusters are composed by nonpure pixels, and less populated clusters are those related to pure pixels. This limitation cannot be avoided even with more sophisticated hierarchical clustering methods. SAD has advantage over Euclidean distance if HAC alone is used. Using HAC and SU in most cases improve the results; in this case, Euclidean distance has advantage over SAD. The downside of using spectral unmixing is that, at least, the number of endmembers is needed, or even endmembers data themselves; this information is not necessarily known and its computation is not straightforward [
Second experiment shows positive correlation between field data entropy and computed entropy almost exclusively using HAC and SU. Given that spectral data comes from vegetation in a forest, high spectral mixing is the most likely scenario; followed by mid spectral mixing in zones having plants with big (greater than spatial resolution) crown surface or dense population. Negative correlations for HAC can be explained as being due to few important clusters containing information of several endmembers, species with very similar spectral signatures or uniform compounds of vegetation. Low correlation in Group 4 can be explained by presence of vegetation with stem radius below the 5 cm. threshold.
A method based on hierarchical agglomerative clustering and spectral unmixing is presented, this method captures spectral variability of vegetation in hyperspectral images; the captured information could be used to assess plant species richness. The method works well even in scenarios with high spectral mixing, but requires a way to determine the number of endmembers or endmembers themselves. Experimental results show positive correlation between the Shannon entropy of the field data and the one computed using the proposed method.
This work was supported by NASA under grant NNX09AV03A and by Laboratory for Applied Remote Sensing and Image Processing, University of Puerto Rico, Mayaguez, Puerto Rico.
The authors declare no conflicts of interest.
Determining number of clusters using critical point “
(
Three instances of “elbow” curves produced during HAC process using real hyperspectral data. Corresponding critical points
(
Entropy correlation using

 

0.1  0.64  0.98  0.95  0.99 
0.2  0.54  0.99  0.92  0.73 
0.3  0.75  0.87  0.83  0.55 
0.4  0.43  0.95  0.39  0.40 
0.5  0.48  0.98  0.67  0.55 
0.6  0.59  0.98  0.58  0.87 
0.7  0.43  0.61  0.50  0.56 
0.8  0.43  0.99  0.07  0.64 
0.9  0.11  0.72  0.31  0.37 
Entropy correlation between field data and proposed method.

 

Group 1  −0.19  0.85  −0.12  0.51 
Group 2  0.59  0.78  0.65  0.87 
Group 3  0.50  0.82  −0.06  0.87 
Group 4  −0.56  0.30  −0.28  0.41 