1.

Introduction

Mapping tree species by remote sensing techniques have been useful in understanding the role of plant species at the landscape scale. Landscape-scale species distributions can be used to help determine classification of land use/land cover, plant response to climate change, detection of invasive species, patterns of plant competition, and spatial distributions of fire fuel loads among other applications.¹ Mapping individual species has been enabled by the advancements in remote sensing technologies, such as hyperspectral imagery or light detection and ranging (LiDAR).

Along with the development of remote sensing devices with the spatial and spectral resolution to map species, there has been a wave of development in statistical and algorithmic approaches to mapping species. We provide a brief review here and highlight that the impact of atmospheric correction has been relegated to a mere preprocessing step although it may have direct impacts on classification accuracy. Colgan et al.² used a two-stage support vector machine (SVM) at both the pixel level and crown level for tree species classification, in which LiDAR measurements were used for crown segmentation. Féret and Asner³ studied the accuracy of various parametric/nonparametric supervised classification techniques and observed that there is a clear advantage in using regularized discriminant analysis, linear discriminant analysis, and SVM. There have been other tree species classification efforts, such as that of Dalponte et al.,⁴ Féret and Asner,⁵ Ghosh et al.,⁶ Immitzer et al.,⁷ Naidoo et al.,⁸ and Ustin et al.,⁹ that share the same approach with minor variations.

Féret and Asner³ compared the classification performance of different hyper- and multispectral sensors, particularly Carnegie Airborne Observatory’s (CAO) hyperspectral alpha system,¹⁰ WorldView-2, and QuickBird. By convolving 72 hyperspectral bands to eight and four multispectral channels available in the WorldView-2 and QuickBird satellite sensors, respectively, they observed that WorldView-2 produced more accurate classification results than either QuickBird or CAO. Clark et al.¹¹ compared leaf, pixel, and crown level measurements to identify important wavelength regions for species discrimination. Although optimal regions of the spectrum for species discrimination varied with scale, near-infrared (NIR) bands were consistently important regions across all scales. Bands in the visible region and shortwave infrared were more important than other bands at pixel and crown scales. Clark and Roberts¹² performed their analysis on higher level data products, such as vegetation indexes, signal derivatives, and signal intensities among others for classification.

Baldeck and Asner¹³ tried to measure beta diversity (turnover of species assemblages between sampling units¹⁴) of different regions using distance measures, such as Euclidean distance and K-means clustering in unsupervised models. Use of these clustering techniques provides a quick assessment of beta diversity, thereby avoiding costly and time-consuming field data collection. However, about 50% of pixels could not be identified to species and were classified as “other,” signaling the need for improvements in techniques, similar argument holds in Baldeck et al.’s later work in Ref. 15. This suggests that there is a certain level in species classification accuracy beyond which state-of-the-art classification techniques cannot discriminate different species, as such are species that are either too rare or too spectrally similar to other species that they cannot be individually distinguished with spectral data alone.

For all of these studies, preprocessing of hyperspectral imagery was required, which can potentially impact the results. The impact of the atmosphere is variable in space and time and usually requires correction for quantitative remote sensing applications. Several researchers have investigated the effects of different atmospheric correction methods for Landsat, QuickBird, and imaging spectrometers on generic land cover detection applications as we elaborate below, but there is little work to address the impact of atmospheric correction for plant species classification. Some approaches to atmospheric correction include scene-derived adjustments, in which in-scene statistics are used, such as the darkest pixel method,¹⁶ or purely empirical methods, where ground-recorded spectral data are required, e.g., the empirical line method.¹⁷ Some involve radiative transfer models, such as the 6S code¹⁸ and moderate spectral resolution atmospheric transmittance (MODTRAN),¹⁹ while others, such as atmospheric correction (ATCOR)²⁰ and fast line-of-sight atmospheric analysis of spectral hypercubes (FLAASH),^21,22 add in situ spectral data to the model.²³ If ground data are not available, radiative transfer models provide a cost- and time-effective solution in atmospheric correction; however, with the availability of field data, ATCOR and FLAASH provide the added value in monitoring the performance and fine tuning the training of the models using field data, hence the focus on ATCOR and FLAASH. Manakos et al.²³ compared the effect of ATCOR and FLAASH atmospheric correction on land cover types in the island of Crete using Worldview-2 satellite data, where FLAASH outperformed ATCOR correction. In particular, ATCOR produced consistently low corrected reflectance values for all targets and all bands.²³ Here, we compare ATCOR versus FLAASH but in the context of tree species classification. The choice of atmospheric correction may be important in species classification because if atmosphere effects that obscure the land surface signal are not properly removed, then separability of different plant species may be diminished. This is particularly important if there is spatial variation in atmospheric effects across the scene to which the species classification will be applied. Also, as sunlight penetration in plants (and their stacks of leaves) is much different to that of nonliving materials (asphalt, gravel, etc), a wider range of signal spectrum is vulnerable to that of atmosphere, which has not been the focus of previous papers.

The data for this study were provided by the National Ecological Observatory Network (NEON) for Ordway-Swisher Biological Station (OSBS) in north-central Florida, United States. NEON includes $\sim 60$ local sites in different ecological domains across the United States.²⁴ Starting in 2017, the NEON remote-sensing airborne observation platform (AOP), carrying a $\text{meter}/\text{sub meter}$ resolution for hyperspectral and LiDAR instruments, will collect images of each site annually for 30 years, thus generating a large amount of data that can be used to map plant species and track their distributions through time. This paper uses prototype image data from a pilot study conducted at Ordway Swisher in advance of the full implementation of NEON in 2017.

The contributions of this paper are as follows. We performed tree species classification using SVM and studied the impact of different atmospheric correction techniques for classification accuracy. We focused on the two commonly used atmospheric correction techniques (ATCOR and FLAASH) that NEON applied to the hyperspectral data collected in 2010. We also explore the use of Gaussian filters for denoising reflectance values. Finally, we examine the use of filters to remove pixels with low vegetation [normalized difference vegetation index (NDVI)] or high shade (NIR), which has been used in a number of species classification studies.^2,8 With this work, we hope to provide guidance for processing the large set of NEON images that will become available starting in 2017 that can be used for species classification.

2.

Data Collection

OSBS covers $37{\mathrm{km}}^2$ in Putnam County in north-central Florida and is managed jointly by the University of Florida and The Nature Conservancy (Fig. 1). OSBS features diverse natural forests and small pine plantations and has a 75-year history of low human impact. The major plant communities in OSBS, as defined by the Florida Natural Areas Inventory, are²⁶ sandhill, xeric hammock, upland mixed forest, baygalls, basin swamp, basin marsh, marsh lake, clastic upland lake, and sandhill upland lakes. The sandhill community is managed using prescribed burning on a scheduled 3-year rotation. The ground sampling for this research focused on a sandhill ecosystem dominated by longleaf pine (Pinus palustris) and turkey oak (Quercus laevis).^26,27

Fig. 1

Location of Ordway-Swisher biological station (OSBS).²⁵

The instrumentation slated for deployment on the NEON AOP remote-sensing payloads in 2017 was not yet available. AOP airborne spectroscopic and LiDAR measurements were performed using existing systems. It is important to note that the actual AOP sensors for NEON starting in 2017 will have better performance, improved conformance of hyperspectral/LiDAR integrations, and better spatial resolution than the sensors used in this study.

Airborne visible/infrared imaging spectrometer (AVIRIS) operated by the Jet Propulsion Laboratory (JPL) onboard on a Twin Otter DeHavilland DHC-6-300 aircraft was used to collect data. Images were collected on two separate days over OSBS: the morning of September 4, 2010 and midday of September 10, 2010 (Fig. 2). Both of the flights were conducted at an approximate altitude of 4000 m AGL at $\sim 90\text{knots}$ with zenith angle of 180.0 and azimuth angle of 0.0 [speed over ground (SOG) $\sim 65$ to 91 knots; NASA JPL AVIRIS flight details on both days^28,29]. The atmospheric conditions were mostly clear with some haze on September 4 and some puffy clouds on September 10. Depending on flight line, pixel sizes ranged from 3.3 to 3.6 m. Hyperspectral data were atmospherically corrected using FLAASH and ATCOR algorithms. Altogether, 8 flight lines and 224 bands were recorded with wavelengths from 365.93 to 2496.24 nm.

Fig. 2

JPL AVIRIS flights over OSBS:²⁷ (a) Flights ground tracks, (b) hyperspectral true-color mosaic, morning September 4, 2010, and (c) hyperspectral true-color mosaic, midday September 10, 2010.

Atmospheric characterization relied on measurements of a CIMEL sun photometer in coordination with the NASA AErosol RObotic NETwork (AERONET).³⁰ Measurements were collected on September 4, 2010 and the derived atmospheric information was used to improve the atmospheric correction of the AVIRIS spectrometer data. Detailed measurements, such as aerosol optical thickness, water vapor, etc., are available online.³¹ NEON personnel performed the orthorectification and atmospheric correction of the reflectance/radiance values, and after extensive studies and ground measurements opted for ATCOR and FLAASH (hence, we focus on ATCOR and FLAASH).

3.

Species Classification

For many pixels in the image, we cannot expect a pure signal of a single species due to the large size of each pixel (3+ by 3+ square meters). Instead, many pixels were linear/nonlinear mixtures of endmembers in each pixel (e.g., multiple canopy crowns, soil, understory vegetation, shadow, etc.). The main species in this study, turkey oak and longleaf pine, have sparse canopies, which increase the amount of soil and branch material visible in the pixels.

3.2.

Signal Preprocessing

The hyperspectral images were loaded in MATLAB using an in-house upgraded version of enviread, initially developed by Dr. I. Howat at Ohio State University.⁴⁴ A check for the consistency of calibration and uniformity of pixel sizes indicated a range of 3.3 to 3.6 M pixel sizes due to various flight and measurement conditions. As different flights have different altitudes and hence pixel resolutions, this is an essential step to take into account. Here, we define a hyperspectral image $I$ with dimensionality $(x,y,w,z)$, where $x\in X=[167000,833000]$ represents the range of UTM easting values, $y\in Y=[0,9400000]$ represents UTM northing values, $w\in W=\{1,\dots ,224\}$ is the index of the reflectance wavelengths, and $z\in Z=\{1,\dots ,60\}$ is the UTM zone of the image. Based on our observations, we take constant $\xi =10000$ as a cutoff point to avoid erroneous sensor readings. There are various types of noise in JPL AVIRIS measurements such as negative reflectance values; the range of hyperspectral reflectance values (range $[-32762,32724]$): one should note that reflectance is the proportion of sun radiance signals, which should be a positive value, but in normalized form, reflectance is between zero and one. The normalization process to remove sensor noisy data is as follows:

Eq. (4)

$$I_{xywz}=\{\begin{array}{ll}0& \text{for}{I}_{xywz}<0\\ 1& \text{for}{I}_{xywz}<\xi \\ \sqrt{\frac{I_{xywz}}{\xi }}& \text{otherwise}\end{array}.$$

For normalization, the negative reflectance values were set to zero and values greater than 10,000 were set to 10,000. Comparing the generated RGB wavelengths to that of the RGB image taken at the same time of the flight, we noticed that the image appears much darker. To adjust the image intensity to a “true color” setting, the square-root of signal returns was used. We excluded wavelengths corresponding to strong water vapor absorption bands in the atmosphere: 1333.2 to 1482.7 nm, 1791.6 to 1967.6 nm, and 2406.9 to 2496.2 nm.

3.2.1.

Impact of low-vegetated/shaded pixels

We tested two filters to obtain pixels that contain greater signal of green vegetation for the canopies. A filter of NIR excludes heavily shaded pixels,² for this data, we set the threshold to 0.33. To obtain pixels with a high signal of green vegetation, a filter of NDVI can be used.⁴⁵ Here, we set the threshold to 0.4 and band 665.6 nm was used for red and 734.1 nm for NIR. However, in our images, we observed that preserving low NDVI/NIR pixels from ground data increased prediction accuracy. Removing low NDVI pixels from tree canopies degraded classification performance by about 10%. This is contrary to the general belief in the literature that removing pixels with low green vegetation contribution (low NDVI) and high shading (low NIR) improves classification performance. Threshold values for NIR and NDVI are highly data dependent (based on sensor type, atmospheric correction, and signal calibration) and were chosen empirically.

3.2.2.

Gaussian filter

The sensor readings were very noisy (Fig. 3). To reduce noise, we take advantage of an abundance of bands (224 AVIRIS bands) and exploit their local-aggregate information by applying a Gaussian filter to reduce cross-band noise. Applying the filter reduced the amount of cross-band noise. Signal transitions are smoother while useful features of the spectrum are preserved. The impact of this preservation in species classification accuracy is demonstrated in our results section. Taking a Gaussian window $w$ of size $N>0$, the coefficients of the Gaussian window are computed as follows:

Eq. (5)

$$w(n)=e^{-\frac{1}{2}{\left(\alpha \frac{n}{N/2}\right)}^2},$$

where $-(N-1)/2\le n\le (N-1)/2$ and $\alpha$ is inversely proportional to the standard deviation ($\sigma$) of a Gaussian random variable ($\sigma =N/2\alpha$). After Gaussian parameters are specified, convolution is performed to apply the smoothing factor. Convolving vectors $u\in {\mathbb{R}}^m$ and $v\in {\mathbb{R}}^n$ gives vector $w\in {\mathbb{R}}^{m+n-1}$, such that

Eq. (6)

$$w(k)=\sum _{j}u(j)v(k-j+1).$$

Fig. 3

Use of Gaussian filter to reduce cross-band sensor noise: water absorption bands are preserved here for displaying purposes.

4.

Results and Discussion

Through extensive analysis, we evaluate the impact of atmospheric correction (FLAASH and ATCOR) on species classification accuracy. Accuracy of species classification was different depending on the atmospheric correction algorithm used and FLAASH atmospheric correction outperforms ATCOR by a margin of about 2% to 4%. First, we evaluate the impact of Gaussian filter; applying a Gaussian filter reduces signal noise with or without presence of water absorption bands (Fig. 4). Our point is not overiterating that removing water absorption bands is a useful step as this has been a de facto approach in the literature; rather, this proves that one can still get improvements on prediction accuracy using proper Gaussian settings even in the presence of highly unusable water absorption bands.

Fig. 4

Impact of Gaussian window on prediction accuracy: (a) before removing water absorption bands and (b) after removing water absorption bands.

Removing low NDVI-NIR pixels from the data set caused a general degradation of performance compared to using all pixels (Fig. 5). In the FLAASH data set, the accuracy is about 70%, and ATCOR yields 66% accuracy. The species in this area have crowns with low leaf density and complex surfaces, such that individual pixels within a crown may have low “greenness” due to wood or soil reflectance through the canopy or canopy shadows from a complex surface. The filters may be removing excessive number of pixels in this instance. The degradation using the NDVI-NIR filter can also be due to the fact that pixel size is large (3 m) and canopies are large (each canopy is from 6 to 107 pixels). This generous marking of canopies includes areas with little greenness (shadows, branches, gravel, etc.; all with low NDVI and low NIR values). Due to the mixing nature of reflectance values, even low NDVI/NIR pixels of a continuous canopy still contain signals from the underlying species, which may not be as green. Figure 5 shows that removing low NDVI/NIR pixels of a continuous canopy actually degrades the performance of the classification model with an impact of about 4%. So, in similar scenarios, where field data consist of large continuous land patches, this observation advises on preserving low NDVI/NIR pixels. It is important to realize that this is only relevant in the context of field data, as the species are already known, and is only useful in training the classifier. In global application of NDVI/NIR filters to entire flight lines, we suggest using the method in the literature; i.e., deletion of low values to remove roads, grass, etc. Here, the benefit of FLAASH over ATCOR can again be observed on average by 2% to 3%.

Fig. 5

Classification results with the removal of low normalized difference vegetation index and near-infrared pixels.

Changing the polynomial degree $P$ impacted prediction accuracy using a polynomial kernel in SVM [Fig. 6(a)]. FLAASH atmospherically corrected data yields 73.5% accuracy, whereas ATCOR results in 69.8%. The simpler the polynomial, the better the performance; a more complicated polynomial leads to a high bias classification model, which performs poorly when evaluated on test data. For FLAASH atmospheric correction ($P\ge 3$), the optimization function does not converge. On the other hand, ATCOR data performs as predicted. Accuracy drops as low as 10.8% and 11.3% with polynomial degrees of 7 and 8 (due to extreme overfitting).

Fig. 6

Parameter tuning for classification algorithms: (a) tuning polynomial order for SVM with polynomial kernel function and (b) tuning $\sigma$ in SVM with radial basis function kernel function.

The radial basis kernel has better performance than a polynomial kernel. As shown in Fig. 6(b), the best results are achieved using FLAASH data, with a peak at 75.3%, while ATCOR comes close at 74.9%. RBF does not show good performance at $\sigma$ values, which are either excessively low or high, where $\sigma$ is the inverse of the width of the RBF kernel (roughly defining the area of influence of a support vector); in other terms, it defines the degree of influence of a single training example. The larger $\sigma$ is, the closer other examples must be to be affected. Because RBF takes data to a higher dimensionality, a small $\sigma$ gives a pointed bump in the higher dimensions, and a large $\sigma$ gives a softer, broader bump. Thus, neither extreme shows a good fit nor a middle point of $\sigma =\mathrm{10,000}$ provides the best results. On the negative side, FLAASH begins with an accuracy of 0% and ATCOR at 19.9%, but they quickly attain a stable region close to each other, while FLAASH demonstrates superior performance in most of the cases. There is a good range of $\sigma$ values ([10, 10000]), which give a plateau in prediction accuracy, implying the RBF kernel has a more robust performance than the polynomial kernel.

Table 3 shows the confusion matrix of the best performing classification model (75.2% accuracy), using the FLAASH correction module and SVM with RBF kernel and Gaussian window length of four. The majority of misclassifications are between pine (other) class and longleaf pine. A similar misclassification can be observed between oak (other) and other types of oak. This was expected because the “other” class contains a mixture of different species of the genera (oak or pine). On the other hand, there is minimal misclassification between the oak versus pine categories. The oak category is rarely misclassified as pine, but pines are misclassified as turkey oak and laurel oak at a low level. There is no misclassification among different oak species. Laurel oak, turkey oak, and live oak are well separable from each other, suggesting that broad-leaf species like oaks can, in general, be well separated.

Table 3

Confusion matrix using radial basis function kernel function with optimal accuracy of 75.2% near peak using the FLAASH correction module.

In our experiments, we observed small but consistently better performance of FLAASH atmospherically corrected data versus ATCOR for tree species classification. This is in conformance with recent observations of Manakos et al.,²³ who found that FLAASH atmospheric correction outperformed ATCOR in endmember classification of land cover types in Crete.

5.

Conclusions

Identification of species using remote sensing technologies, such as hyperspectral and LiDAR sensors, has critical utility in studying impacts of global warming, biomass estimation, and invasive species identification among other issues. In this paper, we report species classification using SVM applied to AVIRIS hyperspectral data available for OSBS in north-central Florida. Performance of the classifier was improved by using a Gaussian filter for denoising reflectance values. We also discuss how incorporating even low NDVI and low NIR pixels can be helpful in improving classification accuracy in some landscapes ground data. Due to the mixing nature of remote sensing hyperspectral data, using such pixels can reduce the bias in maximum margin support vectors in SVM. The images atmospherically corrected using the FLAASH algorithm outperformed the ATCOR algorithm by margins of about 2% to 4%. Our classification model is robust for species classification among different oak species, however, we show minor misclassification between pine and oak species.