Background

Recent decades witnessed a considerable number of alien species being brought into wetland with significant impacts on local ecosystem structure and biodiversity richness across taxa [1]. In particular, variation in hydrological regimes that occur due to natural and anthropic factors strongly affect water-dependent ecosystems [2], [3], [4]. Wetlands are particularly fragile due to host species' high sensitivity to variations in water regime fluctuations. In fact, wetlands respond to nutrient enrichment of associated waters in typical fashion [5]: some shift in plant community composition occurs first after nutrient levels in soil increase, followed by changes in both aquatic and wetland-dependent animal communities [6]. In oligotrophic wetlands, such as peatlands and ombrotrophic bogs, responses to eutrophication may be more rapid, more dramatic, and longer lasting, naturally implying an intrinsic strong correlation between water, soils, and vegetation [7], [5]. Particularly not well understood is the correlation between species richness and rainfall [8], [2]. Yet, an understanding of linkages between soil properties in wetlands and above-ground landscape patterns is critical to developing quantitative soil-landscape models that will aid in detecting, localizing, and characterizing changes in species composition.

Currently, assessment of biodiversity at local and regional scales often relies on fieldwork-based data collection [9]. Species assessment in relatively large or weakly accessible areas is a longtime challenging task for ecologists. As stressed in [10], a key factors need to be determined before a sampling design is ready for implementation, such as: (i) the number of sampling units; (ii) the spatial placement of the sampling units; (iii) clear definition of the statistically meaningful species of concern; and, (iv) an operational definition of a species community. Moreover, field-based approaches are typically labor intensive and costly, and only a small fraction of a study area may be sampled [10]. The same holds true for hydrological and geological fieldworks that aim to better understand the dynamics of water and soils.

The above discussion attests to a crying need for methods that extract information and accurately analyze remote sensing images nowadays available. High-resolution satellite imagery provides detailed spatial characteristics over large areas of ecosystems and offers a promising potential for accurate vegetation mapping [11], [12], [13], [14], [15]. However, most multispectral image classification techniques more commonly focus on spectral discrimination of ground objects for single-species detection [16], [17], and may overlook pertinent information extractable through analysis of spatial, indeed spatiotemporal, pixel intensity variation residing in high resolution images [18], not to mention the multiscale nature of these variations. Noteworthy, the work in [18] emphasizes the need for methods that rapidly and objectively forecast species diversity thru spatiotemporal data analysis. The work in [18] also demonstrates how previous techniques produced poor assessment of species-richness, or diversity, and particularly pairwise species dissimilarity, or diversity. While the work in [19] focuses attention on diversity as an important measure of species dissimilarity between communities, the spatial analysis presented do not consider the time varying aspects of species dissimilarity. Besides the pairwise species dissimilarity, species-turnover, or diversity, reflects the change of environmental variations, such as of rainfall and soil, as dictated by either natural events, anthropic events, or both [2]. Species turnover is a major motivation for the work in [20], [21], and a significant number of research efforts analyze diversity at different spatial scales (e.g. [22], [23], [24], [8], [25], [26]). Yet, apart from [10], time variations in species composition is often overlooked.

Greater Everglades Ecosystem Restoration

In this study we consider the Water Conservation Area 1 (WCA 1) in the Greater Everglades Restoration Area, also known as the Arthur R. Marshall Loxahatchee National Wildlife Refuge. This is a constructed tropical wetland (Figure 1). We demonstrate our approach for 28 year wet and dry seasons using Landsat observations. Among wetlands worldwide, the Greater Everglades Ecosystem Restoration area (GEER) has undergone the most considerable changes in ecohydrological patterns. The changes are due to the constructions of a set of levees and canals aimed to control water flow in the area. Thus, GEER is a reference wetland among ecohydrologists and environmental scientists at large [7], [1], [27], [4], [3]. The work in [4] analyzes the effect of rainfall variation induced by climate change for the whole Everglades National Park. This study evidenced the importance of rainfall for the Everglades, an ecosystem defined as strongly rainfall-dominated. Within GEER, WCA 1 is one area that, in comparison to others, underwent the smallest ecohydrological changes; rainfall thus controls the majority of its water. Thus, WCA 1 is a unique area to test the correlation between climatological and ecohydrological patterns. The choice of GEER to demonstrate our proposed methods is opportune, as there is currently a serious ongoing debate regarding the restoration of Everglades according to the original predrainage patterns; the main concern regards effects that such restoration may bring to the ecosystem [1], [27]. Prior to undertaking such restoration it is very important to develop analytical methods and tools capable of characterizing and localizing multiscale spatiotemporal changes, such as in vegetation patterns, as a function of changes that derive from previous interventions and/or natural changes of climate. Additionally, the analysis of vegetation, plant species-richness in space and time in particular, is also crucial to the development of models that are capable of predicting scenarios of different management alternatives. Management alternatives for WCA 1 are different configurations of the levee system that regulates the ecohydrology in the area. Generally, ecological models are calibrated and validated on observed data, and accurate parameter estimates are therefore critical to enabling these models. Regarding WCA 1, we are only aware of the effort of [28] for evaluating the restoration of the Everglades using satellite imagery. However, [28] did not consider any species richness indicator.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Remote-sensed images for the Arthur R. Marshall Loxahatchee National Wildlife Refuge (WCA 1) during the dry season for the period 1987–2011.

The first three years (1984–1986) images are not represented. The representative region in which the texture analysis is performed is delineated in red for each image. The red regions are characterized by a cloud cover lower than 20%. The green regions identify where the data of species are available. Figure S1 reports the images for the wet season.

https://doi.org/10.1371/journal.pone.0046616.g001

Approach

In a new approach to this problem, we aim to extract species richness in space and time from multispectral satellite imagery, using statistical multiresolution wavelet texture analysis. We adapt the approach in [29], used for retrieval, to texture classification. We demonstrate our analysis of WCA 1's ecohydrological patterns by considering interseasonal and interannual species dissimilarity - which is better known as “species-turnover” [21], [20] - and species richness for twenty-eight years of Landsat observations. The approach consists of applying wavelet decomposition, and statistically modeling then comparing these coefficients either thru a non-parametric histogram, or thru an appropriate probability density function. We elect a model that embraces a wide spectrum of central and tail behavior that can be represented by varying just two parameters at each subband. Specifically, the coefficients from each scale, even each subband of the decomposition are considered samples from a generalized Gaussian (GG) probability density function (pdf). The coefficients provide estimate of the characterizing GG pdf parameters. We next assume independence across scales and across subbands within scales and obtain a joint pdf for each textured image region. The independence assumption is only an approximation, but experience has shown that it provides accurate results when the objective is to quantitatively examine differences across regions and time points. There are many ways to quantitatively assess differences between pdf's; we choose the relative entropy, known as the Kullback-Leibler divergence (KL). Other methods, such as likelihood ratios, and various metrics, could also be examined. We use the KL divergence and the Shannon entropy to assess species dissimilarity in time ( diversity), and species richness ( diversity) respectively. Entropy is widely reported as a proxy of species richness [30], while spectral heterogeneity measured by reflectivity is a measure to quantify the entropy [31], [32].

We validate our predictions on independent data from the Global Biodiversity Information Facility [33] that are composed by field-data. We do not compare directly our method with other multispectral analysis methods (e.g. see [18] for a review) as we prefer to provide a novel theoretical and methodological framework - tested against long duration data records - for analyzing multispectral images.

Previous Research

To the best of our knowledge this is the first time that a statistical model based multiresolution texture analysis is applied to satellite imagery for quantifying species-turnover, a multiscale spatiotemporal phenomenon. Previously, the work in [19], [34] and recently [35] used the KL divergence for estimating diversity in space as a theoretical construct without any model. In [36], texture analysis predicts avian biodiversity richness based on satellite imagery. Species turnover, however, was not considered. Other previous texture-based calculation used the intensity of imagery for estimating biodiversity variables. Texture-based analysis have investigated species habitat relationships [37], successional attributes of forest species [38], classification of species [39], species richness as a function of spatial and spectral resolution [15], intercorrelations of vegetation biodiversity and soil dynamics at both local and regional scales [40], and different band combinations of images [41].

None of the above authors, however, specifically account for multiple scale variations in both space and time. Statistical multiresolution texture analysis proved successful in other applications; it is previously applied to classify stem cell colonies [42], [43], [44], thus enabling non-invasive analysis of these cells. Non-invasive analysis of cells is a disruptive technology that carries the potential of supplementing, even replacing invasive and at times destructive chemical biomarkers. As such, in addition to lowering cost, it provides two advantages: (1) it provides a quantitative statistical assessment of these cells; and (2) enables the preservation of these cells for their intended use, such as drug testing and tissue formation. For the ecological study of concern here, the analysis will have similar advantages: it will reduce the need for costly field fieldwork-based data collection [9]. It is also non-invasive, while fieldwork data may also perturb the ecosystem and the reliability of the collected data may be low because of the generated disturbance or because of the limitedness of the sampled data [45], [1].

Summary of Contributions

The overall objective of this study is to investigate the spatiotemporal relationships between soil, water, and vegetation patterns derived from remote sensing data in a tropical wetland. We adapt and apply a multiresolution statistical texture analysis to remote sensing imagery. The methodology developed aims to account for both the statistical and multiscale spatiotemporal nature of the above mentioned relationships, thus enabling more parsimonious inference at higher spatial and temporal resolution. We also investigate the accuracy of the KL divergence as a measure of diversity in time versus conventional approaches such as the difference of the Shannon entropy. We detect: (i) the hydrological footprints in term of correlation between water, vegetation, and soil changes; (ii) the interseasonal variation of species-richness in which dry and wet seasons are compared for each year (Figure 2); and, (ii) the interannual variation of species-richness for the same season between consecutive and non-consecutive years (Figure 2). We verify that the use of the KL divergence estimated on texture performs better than conventional methods. This analysis is performed successfully on low-medium resolution data (Landsat images). Moreover, we illustrate the results with a novel color based visualization that provides insight into the ecological variations, and enables a quick recognition of anomalous data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Graphical explanation of the analysis performed for the WCA 1.

Landsat images (from 1984 to 2011) are acquired for the dry and the wet season. The example is reported for the years 1997 and 1998. The changes in vegetation composition are analyzed using and diversity among seasons and among years. For the interseasonal analysis the time-scale is on average six months between seasons of the same year, while for the interannual analysis the time-scale is about a year between the same season of different years. and diversity from data are compared to the estimates of the Shannon entropy (Equation 5) and of the KL divergence (Equation 4) for the green band of the images respectively.

https://doi.org/10.1371/journal.pone.0046616.g002

Satellite Imagery

The satellite images of the Water Conservation Area 1, described in Figure 1 and Text S1, are derived from the Landsat database of [46] and [47]. Figure 1 and Figure S1 show WCA 1 for the 1987–2011 period in the dry and wet season respectively. The resolution of these images is 30 m. The images from 1984 to 1998 are from the L4–5 TM dataset, from 1999 to 2003 from the L7 ETM+ with SLC-on (1999–2003), and from 2003 to present from the L7 ETM+ with SLC-off. The Scan Line Corrector (SLC) stopped working on May 31, 2003 and that caused striping of remote sensed images [48]. Thus, we do not consider in our analysis the striped parts of these images. Each Landsat image is cropped along the boundaries of the Water Conservation Area 1 provided by the South Florida Water Management District. The representative area of analysis for each image is selected in oder to guarantee a cloud cover lower than 20% on average for each selected area and a minimum area requirement for the texture method (Section). At the same time, we strived to maximize the extension of the selected area and its representativeness of the whole landscape. For each year two images are collected, one for the dry season, December to April, and one for the wet season, May to November. Images were selected in January and August, the most representative months for the dry and wet season respectively. In Figures 1 and S1, the red squares show the regions selected for texture analysis. We analyze the satellite images for each visible band defined as:

We analyze the satellite images for each visible band defined as:

• Band 1: (Blue; electromagnetic wavelength: 0.45–0.52 ), which is useful for mapping water, differentiating between soil and plants, and identifying manmade objects such as roads and buildings;
• Band 2: (Green; electromagnetic wavelength: 0.52–0.60 ), which spans the region between the blue and red chlorophyll absorption bands, and shows the green reflectance of healthy vegetation (thus vegetation that changes across seasons and years). It is useful for differentiating between types of plants, determining the health of plants, and identifying manmade objects;
• Band 3: (Red; electromagnetic wavelength: 0.63–0.69 ), which is one of the most important bands for discriminating among different kinds of vegetation. It is also useful for mapping soil type boundaries and geological formation boundaries. It is considered as a proxy of soil types.

Species Occurrences

The species richness for WCA 1 is compiled by assembling together data from fieldworks of [45], the Global Biodiversity Information Facility database [33], and the Comprehensive Everglades Restoration Project database [49]. In order to compare the predicted species richness from the analysis of the green band of Landsat images, we build a species richness matrix at 30 m resolution from the aformentioned sources. For more information about WCA 1 we refer the reader to Text S1.

The “base” local species-richness (or diversity) of plant communities is built from the Global Biodiversity Information Facility [33] data that provide occurrences of species in space and time at 1 resolution. The green squares in Figure 1 and S1 indicate the regions where the GBIF data are available. We consider data from 1984 to 2011 such as for the Landsat images. For the GBIF data we downscale the information of species richness from 1 to 30 using a simple coarse-graining algorithm [50]. A grid of smaller boxes is created and the number of species are counted at the resolution of the smaller grid. The estimation of species richness is also refined with point data of plant species occurrences from CERPzone, that is the database of the Comprehensive Everglades Restoration Project [49]. CERPzone data are available from 2001. In the species richness matrix we also include the data of [45], where a total of 30 plant species are found along the Loxahatchee National Wildlife Refuge transect which goes from the westernmost to the easternmost boundary of WCA 1 at its maximum width. According to [45] the species-richness at a given site never exceeded 8 . Of these 30 species, only 11 were found in both 1989 and 1999 (Table 2 in [45]).

The occurrences of these three datasets are compared together after creating a grid over WCA 1 in which the unitary pixel is characterized by an area of 30 . The point occurrences of CERP and [45] are downscaled to 30 following the approach of [51] that is a shot-noise Cox process method. All the unique occurrences of species from all datasets are assigned to each pixel according to a distance criterion. Each occurrence characterized by geographic coordinates is assigned to the nearest neighbor pixel. The local species-richness or diversity is calculated by the sum of unique species in each pixel, and the average local species-richness of WCA 1 is the average over all the species-richness values of each pixel. Species turnover or diversity is calculated on the data of [33] as complementary to the Jaccard Similarity Index (JSI) evaluated in time at resolution of 30 m. JSI is given by the ratio of the number of common species in two pixels considered and the number of all species in both pixels. Specifically, , where and are the numbers of species present in pixel and at different seasons or years). We also note that diversity is the number of distinct species in the whole domain of the green squares (Figure 1).

Specific Study Hypothesis

Our study is based on the following hypothesis formulated on some previous studies and data.

1. Rainfall is the main driver of vegetation patterns in WCA 1 and in tropical wetlands [1]. This is evidenced by hydrologic studies [4] and by analysis of satellite imagery pre- and post-construction of the levee-canal drainage systems [1] We do not anticipate any relationship between changes in soil, water, and vegetation richness in WCA 1 because non-linearities were evidenced among these variables [5]. However, because of the high seasonality of climate [8], [1] we expect variation of species richness within each year for the wet and dry season such as was reported for single species [52].
2. Landsat RGB imagery, despite its low resolution,reveals information about the spatiotemporal structure of ecosystem components [53], [54]. As stated by [54] high resolution imagery is not always available for free for all the regions in the world. Moreover, imagery of higher quality does not always guarantee better estimation of biodiversity variables. This is particularly true in tropical ecosystems where the resolution of high quality imagery can be too small compared to the extent of each single species [54]. We believe that the improvement of methodologies for assessing biodiversity variables and other environmental variables of ecosystems such as in this paper, can overcome the low quality of imagery such as the Landsat imagery.
3. The Shannon entropy of the green band can be considered as the spectral heterogeneity (reflectance) of plant species in water-dominated ecosystems [55], [56], [57], [58], [59], [60], [61], [62], [63], [64], [65], [66]. The blue and red bands can be considered as the spectral signatures of water and soil heterogeneities in the ecosystem [55]. Different species have different reflectance levels for different bands of the Landsat images. The higher the range of reflectivity, the higher the entropy because there are potentially more plant species with different degree of reflectivity. In ecology there is an large discussion that different reflectance levels do not always characterize different species, but “functional species” [55], [58], [59], [67], [53], [68]. The issue about species diversity versus functional diversity is a longstanding issue in ecology. Functional diversity, which is different but related to taxonomic diversity (i.e., species richness), is crucial to ecosystems and the services they provide to humans [59]. In this paper we talk about “species” rather than “functional species”; nonetheless, we make clear that similar reflectance levels may characterize functional species or vegetation types (within the same spectral group) rather than the same species. No distinction is made here between invasive and endemic species; however, the variation of species richness can also be attributed to the invasion of non-native species and inform species management. Other sources of variation in surface reflectance, such as directional effects and shadowing, are also sources of texture variations in Landsat scenes. Nonetheless, we believe our algorithm is capable of detecting the average texture generated by species differences.
4. and diversity are related to the Shannon entropy and to the KL divergence of the green band respectively. Previous studies investigated these relationships based on data and theoretical constructs [19], [35]. We also hypothesize that and diversity in time are independent measures of biodiversity richness as suggested by previous studies [69], [70], [71], [72], [73], [20]. We conjecture that the average diversity is a lower but close estimate of diversity. This is because the distribution of species is quite homogeneous within WCA 1 as verified by data [33] and as reported in [74]. Thus, because the number of species is on average the same in any representative region within the whole ecosystem and for the period considered, scale-invariance of diversity is hypothesized.
5. The difference between Shannon entropies provides a worse estimate of diversity in time than the KL divergence. We expected that this occurs because the KL divergence, that is not calculated on the intensity of the Shannon entropy, potentially accounts for the pairwise interactions between vegetation communities in time (i.e. the mutual information in information theory). On the contrary, the difference between Shannon entropies does not capture the community pairwise interactions [19].

Texture Image Analysis

We model the region of interest as textures. Texture models abound in the literature and textbooks (see e.g. [75]). We favor a model that captures what in our view are two essential properties of texture: variations of intensity across pixels that: (i) occur at multiple scale, and (ii) are stochastic in nature. Such a model suggests that a wavelet analysis can decompose the spatial intensity variation into multiple scales, and subbands within scales, after which the coefficients of the decomposition at each scale can be analyzed statistically. The across scale statistical analysis characterizes the texture. This is the approach followed in [29] for texture retrieval. As wavelet analysis decomposes a signal locally according to orientation and scale, it is especially apt for modeling texture, characterized by intensity randomness at multiple scales. Note that perceived image texture is closely related to the degree of random fluctuation in image gray-scale intensity at multiple frequencies. An -level wavelet decomposition produces detail subbands, three per level, whose coefficients convey information about the fluctuation at a particular scale and orientation (one oriented horizontally, one vertically, and one diagonally). We compute textural features from these subbands by modeling the empirical pdf of the coefficients in each subband. Moving from one level to the next the scale increases. The above method was also adapted to texture comparison and classification, and used to non-invasively and quantitatively classify stem cell colonies [42], [43], without using chemical markers, thus preserving the colonies for use. The method is also used to classify nuclei in [44], [76], [77]. We will adopt this method here to compare Landsat images of regions across seasons and years, and demonstrate in the next section that it successfully predicts species turnover. A complete description of this approach is beyond the scope of this paper, byt the approach is thoroughly developed in [29] and [42]. We briefly outline the steps below. For representing a texture by a probability model:

Ecological Equivalence

The ecological equivalence of texture analysis is based on the assumption that the entropy and the KL divergence represent the potential species-richness and the dissimilarity in time of species-richness within communities ( and diversity respectively). In the following we propose a theoretical description of entropy and KL divergence in ecology making analogies to the definition of these quantities in image analysis (Section). Because of our hypothesis on the reflectance of species, the entropy in shades of green means diversity in types of plants, which determines diversity. KL divergence between two regions at different times means different shades of green for the two regions, which determines diversity in time. and diversities are among the most employed theoretical concepts in ecology and biodiversity conservation. For most ecologists, diversity traditionally reflects the within-habitat diversity [80], whereas diversity is the component of “total diversity” that is produced by differences in species composition among the sampling units [81]. The need to partition diversity within and among habitats has both theoretical and applied interests. Species-richness is a diversity index of order zero, Shannon entropy is a diversity index of order one, and all divergence measures are diversity indices of order two [30], [18]. Because the Shannon entropy is a first order measure of species diversity [30], the variation of Shannon entropy in time is as well a first order measure of species diversity between the same or different regions.

The Hill order [82] is the order of any information measure that is derived from the generalized expression or limits of such functions as approaches unity, where is a probability distribution function and is the number of equally common species that compose a community. The diversity metrics from are often called “Hill numbers”, but they are more general than [82]'s derivation suggests [30]. The exponent and superscript may be called the “order” of the diversity; the true diversity depends only on the value of and the species frequencies, and not on the functional form of the index. For that approaches unity the information functions are the species richness, Shannon entropy, all Simpson measures, all Renyi entropies, all HCDT or “Tsallis” entropies, and many others. All values of less than unity give diversities that disproportionately favor rare species, while all values of greater than unity disproportionately favor the most common species [83], [30].

Generally the higher the metric's order the better the estimation of the ecological metric. With the knowledge of entropy the species-turnover could be potentially calculated as difference of entropies between two different time steps. However, we consider whether the KL divergence is a better measure with respect to the difference of entropies. In fact, the KL divergence captures the pairwise variation of species-richness of local communities and not only their independent variation of species-richness [19]. For example [84] showed how the pairwise interaction is not negligible in grasslands and this has profound implications for the ecosystem functioning. Recently the KL divergence was related to the Shannon entropy and the species-richness in the whole ecosystem analyzed [35]. We consider that the species-richness and dissimilarity estimations are representative of the whole region extending the analysis from the subregions in which the texture analysis is performed. The entropy calculated by our texture analysis method can be assumed to be an estimate of the maximum likelihood estimation of the Shannon's index [19] defined as:(5)where is the observed sample fraction a local community (, where can be considered as the number of individuals of species , and is the number of individuals of the -th species) at time . The entropy here introduced has the same meaning of the entropy calculated using the intensity of the green band images. The same entropy can be calculated over an other local community for an other function . The entropy as defined above is one of the most commonly used index of diversity [19], [71], that is the number of species in a subarea of the region considered, in theoretical ecology and ecological modeling. Certainly there are other metrics to estimate diversity, but the entropy is one of the most profound and useful of all diversity indices. However its value gives the uncertainty rather than the diversity. Thus, the entropy needs to be rescaled to diversity given some data or calculated as in [30]. The entropy is easily derived from the “order generating function” of [81]. The predicted diversity is tested vs. the estimation of diversity from the data described in Section at resolution of 30 m.

The KL divergence calculated by our texture analysis method is an estimate of the divergence:(6)where is the discriminant information introduced by [85] (Equation 3). Thus, KL is the symmetric version of the divergence 4. Equation 6 is estimated using Equation 4. Let`s assume that local communities (or regions) and are composed by the same species (labelled from 1 to ) with probability distributions and respectively, with , and , . and are evaluated at different time step, and where in our analysis is approximatively six-month and a year for the interseasonal and interannual comparison of the species dissimilarity, respectively. and are the Shannon entropies of communities and calculated as in Equation 5, and is the “reciprocal information”. The reciprocal information can be assessed only after the knowledge of KL (Equation 4) and of and . The reciprocal information represents the pairwise interactions among vegetation species in different communities in time (or in space) which is not captured by completely by the differences of entropies of the two local communities. The KL divergence gives a measure of how much the distribution of species within the communities differs from one another. This has been traditionally called diversity in ecology. The diversity is the most useful metric from which it is possible to assess the number of species and their abundance. We do not use the KL divergence for measuring the dissimilarity of species in space within the same region or between different regions although this is a possible operation that can be performed. Thus, Equation 4 is a proxy of Equation 6 of the landscape analyzed.

Conclusions

We analyze species richness, species turnover, and other spectral heterogeneities derived from satellite imagery for a constructed wetland ecosystem. We use wavelet-based statistical multiresolution texture analysis, a method that, to the best of our knowledge, is new to the field. This method accurately estimates parameters for the marginal distribution of wavelet coefficients using Generalized Gaussian density (GGD) and provides a closed form expression form for the KL divergence between GGD models of different textures. We demonstrate the method on remotely sensed images of the Water Conservation Area 1 in the Greater Everglades Ecosystem Restoration in South Florida but it can applied to any ecosystem. The results suggest that the method is indeed promising for the analysis of species-richness of any ecosystem with high spatial heterogeneity. The method is shown to provide better estimates of and diversity in ecosystems where data-sampling is not feasible (e.g. inaccessible regions) or when it is partially available in time. In particular, statistical wavelet based multiresolution analysis provides a means for estimating divergence between two textures, specifically the Kullback-Leibler divergence between the pair of density functions representing the textures. The KL divergence proved to be a near perfect predictor (R² = 0.98) of species turnover, or diversity. Additionally, the visualization representing the KL divergence results between texture pairs provide quick insight into species turnover across many years and seasons. It also enables quick recognition of anomalous data. Texture modeling is also helpful to the theoretical understanding of fundamental ecosystem processes, classification of land-cover, classification of species [39], and as inputs to ecohydrological models that are capable of predicting hydroperiod and runoff of wetlands. We anticipate further effort in using Generalized Gaussian, as well as other non-Gaussian multiresolution texture methods, together with the KL and other divergences for comparing textures, in order to achieve several tasks of importance, more accurately at a higher spatiotemporal resolution. These tasks include detecting single species occurrence and abundance through segmentation analysis citref [76], assessing species-dissimilarity in space, and more generally, analyzing satellite imagery including stereoscopic images. Mathematical features for representing textures other than wavelet coefficients and their distribution could also be of interest, especially if they provide models that are simultaneously simple and parsimonous. Moreover, the application of the presented model to other biological systems (for example for the identification of cell communities [44], [76]) and other scales of analysis of ecosystems is a promising research direction. Finally, the methods reduce the need for field-work, while enabling more effective, less costly monitoring, inference, and decision making support.