Another remote sensing option, related to the assessment of biomass, is the combined use of different wavelengths of light to differentiate tree species in stands, i. e., enabling biomass-per-species maps. This is relevant because biomass functions for total biomass and biomass components are usually species-specific (Chap. 3). So knowing the species composition of stands would surely improve the prediction of total biomass. This species identification was only made possible with the devel­opment of spectral resolution increases from multispectral to hyperspectral sensors (see Fig. 2.7). Hyperspectral sensing, also called imaging spectroscopy, samples more regions of the electromagnetic spectrum than multispectral sensors, such that absorption from specific leaf pigments, canopy structure, or leaf water content can be estimated (Curran 1989; Wessman et al. 1989; Yoder and Pettigrew-Cosby 1995;

Kokaly and Clark 1999; Kokaly 2001). The electromagnetic canopy reflectance signature provides enough detailed data to discriminate between signatures of different species. It is therefore possible to apply species specific models to stratify a community into more representative components and calculating biomass by component using for example canopy structure profiles from LiDAR (Chambers et al. 2007). The reader is referred to seminal studies on this topic, including Gong et al. (1997), Martin et al. (1998), Fung et al. (1999), and van Aardt and Wynne (2001), with an extension to operational airborne data (van Aardt and Wynne 2007) and commercial plantations (van Aardt and Norris-Rogers 2008). These approaches also have been borne out in savannah regions by Cho et al. (2010). These studies report classification accuracies >90 % for deciduous species and as high as 85 % for coniferous species, while the commercial species, specifically Eucalyptus sp., have proven to be spectral separable with accuracies at approximately 90 %. All of these studies have approached the challenge by subsetting the hyperspectral data to those wavelengths necessary for separation of a specific set of species. In other words, an operational workflow should be preceded by a pilot study: (i) identify the species of interest, (ii) acquire hyperspectral data, (iii) determine which wavelengths are necessary to separate the species on a spectral basis, and (iv) assess the classification accuracy. And herein lies the caveat — hyperspectral data are by design “oversampled” data, i. e., we have more data than we need.

We therefore need to subset the data to the wavelengths required for a specific application in order to develop statistically robust models, or models with a reasonable number of independent or explanatory variables. An example may be warranted: Imagine that a pilot study shows that wavelengths at 452 nm (blue), 622 nm (red), 1,050 nm (near-infrared), 1,452 nm (shortwave-infrared), and 2,248 nm (shortwave-infrared) are essential to separating three coniferous species at an accuracy of 85 %. The logical approach, for an operational implementation, would be to approach a company that specializes in acquiring imagery via air­borne detector/s that can be “programmed” to these wavelengths, via the use of wavelength-specific filters. Such companies and sensors do exist, but they typically only operate silicon-based sensors, which are sensitive to the wavelength range of roughly 380-900 nm; the conundrum is obvious: in order to acquire the necessary wavelengths for the species classification application, one would have to build a relatively expensive sensor, unless the subset of wavelengths can be constrained to the silicon range. This latter solution often is viable, but may come at the cost of slight decreases in accuracy. However, this approach is actually not infeasible, given the lower cost and higher prevalence of silicon sensors; many forestry applications, e. g., species classification, nutrient mapping, moisture stress detection, etc., could be constrained to this wavelength range and thus executed on an operational basis. In fact, leaf-level studies have shown extension to airborne cases, thus adding to potential operational implementation.

Results from a study in tropical forests of Costa Rica indicate that there are spectral differences among species that permit classification at leaf to crown scales (Clark et al. 2005), further corroborated by van Aardt and Wynne (2007) in a mixed oak-pine forest in the Virginia piedmont, USA. However there are also temporal, spatial, and spectral variation within populations and even single individuals of forest tree species that will inevitably decrease classification accuracy and need to be assessed on a as-needed basis. A major challenge is to develop classification schemes that can maximize the spectral, spatial, and temporal information content of digital imagery while accommodating inherent variation within species.