Hyperspectral remote sensing presents the opportunity to collect narrow contiguous wavebands numbering over hundreds across the electromagnetic spectrum. These valuable information can be related to physical and chemical composition of the material in the image scene. Compared to multispectral imagery, hyperspectral remote sensing is rapidly emerging as a promising technology with the potential to improve classification performance of mangrove species. Nevertheless, the development of methods to harness the wealth of information available in the hyperspectral data for mangrove classification is still at the initial stage. Continued development of new approaches to harness the valuable information in hyperspectral imagery is therefore encouraged. Most traditional methods often employed for mangrove classification (e.g., pixel-based classification, machine learning, etc.) utilize spectral characteristics (both at canopy and leaf levels) as well textural features from hyperspectral imagery. Recently our lab (ARSL) proposed spectral complexity metric as a new approach to analyze hyperspectral data for mangrove classification. The study utilized full range airborne hyperspectral data (Visible-Shortwave Infrared) over Sierpe Mangroves in Costa Rica collected as part of the MAC13 project to quantify spectral complexity for mangrove classification. In this study it was found that spectral complexity metrics produced spectral signatures that can be associated with mangrove and terrestrial forest. These spectral signatures can be used to not only differentiate both ecosystems but also for delineating mangrove species.

From (Osei Darko et al., 2021). A) Ground photograph from the Sierpe mangroves, (B) Subset of the CARTA 2005 aerial photographs from the mangroves, (C) Subset of a CASI 1500 HSI (R: 748 nm G: 550 nm B: 519 nm) flight line acquired in 2013 illustrating the same subset of mangroves as seen in B. The delimitation of terrestrial forest and mangroves is more apparent from the HSI than from the aerial photograph.

What is Spectral Complexity

Spectral Complexity metrics are two well-known measures in statistical physics namely Mean Information Gain and Marginal Entropy. When they are adapted to reflectance, they describe the spatial (Mean Information Gain) and aspatial (Marginal Entropy) heterogeneity of the spectral information. One key difference between the two information theory measures has to do with the way they are computed. For instance, the computation of Mean Information Gain takes into account the spatial variation (neighborhood of pixels) in reflected radiation across the landscape to provide an estimate of reflectance spatial heterogeneity whereas marginal entropy provides an estimate of the aspatial spectral heterogeneity (i.e., the probability of locating a pixel independent of its location in the image). The values of the spectral complexity are an indication of the variability of the reflectance. A value of 0.4 is generally accepted as the threshold between low and high spectral complexity From (Osei Darko et al., 2021) Comparison of changing MIG (0.2–0.7) with a constant ME. Reprinted from Ecological Indicators, 8/270–84, Raphaël Proulx, Lael Parrott, ‘Measures of structural complexity in digital images for monitoring the ecological signature of an old-growth forest ecosystem’, page 15, Copyright (2008), with permission from Elsevier

How is Spectral Complexity Computed?

The implementation of SCM is based on a non-overlapping window (i.e., edges of a square blocks or tiles of imagery) approach. Prior to the computation of spectral complexity, a key consideration will be to determine the optimal window size necessary for the computation. For instance, in our study mentioned above, we explored different window sizes to ensure appropriate pixel pairs are captured that could depict both the spatial and aspatial heterogeneity within a unit area. Overall we tested three window sizes, starting from a 25 x 25-pixel window, which represents 0.4 ha, to 100 x 100 pixels (6.25 ha) and 200 x 200 pixels (25 ha) for mangrove extent mapping. To discriminate mangrove species, we examined two additional smaller windows sizes 10 x 10 pixels (0.063 ha) and 5 x 5 pixels (0.016 ha). For the above study, since the predominant mangrove species in the study area occur in stands > 1 ha (40 x 40 pixels), the above-mentioned window sizes were deemed suitable for species classification.

From (Osei Darko et al., 2021) Schematic diagram showing the non-overlapping window approach adopted in the implementation of spectral complexity metrics on hyperspectral imagery. The non-overlapping window outlines are not drawn to scale. Red dashed lines show an active processing window with a diagram showing an example of the MIG (k = 2) computation reprinted from Ecological Indicators, 9/1248–1256, Raphaël Proulx, Lael Parrott, Structural complexity in digital images as an ecological indicator for monitoring forest dynamics across scale, space and time, page 1251, Copyright (2009), with permission from Elsevier

Can Spectral Complexity Metrics improve the classification performance of Forest and mangrove extent?

When spectral complexity metrics and reflectance were used to train multiple classifiers, we found that while the highest accuracy separating mangroves from terrestrial forest in the VNIR was found for the reflectance data at the native spectral resolution (OA = 98.8%), classification accuracy of both spectral complexity metrics outperformed the original spectral reflectance in the SWIR by 2.7%.

From (Osei Darko et al., 2021) The mean classification accuracy for mangrove and forest achieved for each of the input datasets assessed in this study (i.e., Reflectance, MIG, and ME) for both VNIR and SWIR imagery. The error bars illustrate the minimum and maximum overall classification accuracy achieved for each dataset.

In comparison to reflectance can spectral complexity improve mangrove species separability?

When spectral complexity and reflectance were used to train multiple classifiers in the above study for mangrove species discrimination, it was found that unlike the SWIR reflectance, a transformation of VNIR reflectance to a measure of spectral complexity (25 m window size) can improve the classification performance (i.e., up to 93.6%) compared to the accuracy attained from the VNIR reflectance at the native spatial resolution (89.7%).

From (Osei Darko et al., 2021) A comparison of the mean classification accuracies for different window sizes 62.5 m, 25 m, and 12.5 m of reflectance data and spectral complexity. The error bars illustrate the minimum and maximum overall classification accuracy achieved for each dataset.

Limitation of Spectral Complexity Metrics ?

The key limitation of the spectral complexity is the spatial resolution of the final product, equivalent to the size of the non-overlapping window used for the computation. This implies that mangrove communities smaller than the spatial extent of the window size cannot be detected. Based on our findings above, we propose that future studies may take advantage of the ultrafine spatial resolution (<5 cm) associated with the rapidly evolving UAV hyperspectral systems to quantify spectral complexity to a finer spatial scale (e.g., <4 m) necessary for the detection of mangrove species at the tree level. Additionally, the results from our study suggest that spectral complexity can be implemented on hyperspectral imagery from current and forthcoming spaceborne systems to permit monitoring of temporal trends in mangrove extent and species at large geographic scales (i.e., landscape, national, and global scales).

Mentioned Study

OSEI DARKO, P., KALACSKA, M., ARROYO-MORA, J. P. & FAGAN, M. E. 2021. Spectral Complexity of Hyperspectral Images: A New Approach for Mangrove Classification. Remote Sensing, 13, 2604.