2.1.

Study Site

UKL in Oregon, and (Fig. 1) consistently experiences cyanobacterial blooms each summer. The shallow lake (3-m average depth) has become increasingly eutrophic due to human activity, such as the draining of wetlands for agriculture, which has resulted in higher nutrient inputs.¹³ Cyanobacteria blooms are so persistent and extensive that an industry has developed around one of the lake’s most common genera, A. flos-aquae, which is harvested from the lake and sold as a health supplement. Although some strains of Aphanizomenon have been shown to produce toxins, the A. flos-aquae has not been found to be toxic in UKL.¹⁴ However, Microcystis is also consistently present in the lake and does produce microcystins in this environment, which is a problem for both the harvesting of Aphanizomenon as well as the recreational activities on the lake. Blooms in this region often proceed from a progression of dominance by Aphanizomenon, followed by mixed blooms, and then by dominance of Microcystis, sometimes with multiple cycles in one season and typically in mid- to late-summer.¹¹ The prolonged appearance of these blooms is also thought to affect the fish populations in the lake, particularly two species of endangered suckers, because the oxygen demand of decaying blooms can lead to hypoxic conditions.¹³

Fig. 1

Location of UKL, Oregon.

The same four general locations on UKL were sampled in summer 2016, from July 15, 2016 (week 1) to September 15, 2016 (week 10). Raw water samples were collected in opaque plastic bottles, chilled and sent overnight to NIST for spectral analysis. Samples were processed as soon as they were received or refrigerated until they could be processed. Wet slides were prepared with algae taken from the top of each sample bottle. A reference slide was prepared with distilled water and cover glass.

2.2.

Hyperspectral Microscope

The NIST hyperspectral microscope consists of two parts: the microscope component is an Olympus MVX-10 microscope with an MX-2X Objective and a 1 to $4\times$ magnification changer, and the hyperspectral imaging component is a Surface Optics SOC-710 hyperspectral camera (Surface Optics, San Diego, California). The SOC-710 measures 128 spectral bands, at 4.69 nm spectral sampling over the 400 to 1000 nm spectral range. The finest spatial resolution of the system is on the order of $<0.2\mu \mathrm{m}/\mathrm{pixel}$. Illumination was provided by three quartz-halogen lamps set with dichroic reflectors to match the solar spectrum with a high color temperature.

Before sample measurement, reference data cubes were collected from the reference slide placed on top of a Spectralon^® disk (Labsphere, North Sutton, New Hampshire), and integration times were optimized for each desired magnification level. Data cubes were then collected from the sample slides, generally one slide per sample bottle. Depending on integration time, image acquisition took between 2 and 15 min to complete. Two to 10 images were collected per slide, depending on algal abundance. A fresh reference slide was measured after data collection was finished, 4 to 6 h later. Slides were kept out of direct light when not being measured. Figure 2 shows examples of the true color composites from the hyperspectral microscope images.

Fig. 2

True color composites derived from data cubes collected by the hyperspectral microscope. (a) Aphanizomenon, Microcystis, and green algae. (b) Aphanizomenon at full magnification, with a visible heterocyte. (c) Microcystis at full magnification, with green algae (circled). Scale bars are 50, 20, and $20\mu \mathrm{m}$, respectively.

2.3.

Data Analysis

Data cubes were normalized in ENVI (Exelis, Boulder, Colorado) to the reference data cube. A principal component (PC) transformation was applied to a spectral subset of bands 7 to 65 (402.5 to 701.1 nm) to avoid the noisy regions of the spectrum. The spectral derivative of each cube was calculated with the ENVI user function “Spectral_Derivative.” The derivative transformation was applied to a subset of bands 13 to 59 (432.6 to 669.5 nm) to encompass the main features that were observed to differ between the derivative spectra of Microcystis and Aphanizomenon. Basic statistical and hyperspectral processing techniques were utilized to determine the separability of the spectral classes in this and many other applications. Both principal components analysis (PCA) and derivative transformations help determine the separability of basic algal signatures.

The normalized data, the PC data, and the derivative data were then analyzed with several of the algorithms and classification methods included in ENVI to evaluate the separability of Microcystis and Aphanizomenon. PCA was used to determine quickly whether different genera were dissimilar enough to classify separately. These transformed datasets were chosen as basic hyperspectral processing techniques that are used to determine if specific spectral signatures are separable in a dataset.

Analysis of the normalized and derivative spectra was carried out using the ENVI linear spectral unmixing algorithm. Endmembers were collected from the input image as representative cells and example images from the week of August 8 (week 5) were chosen to extract representative spectra for Aphanizomenon, Microcystis, and green algae because they represented outstanding visual examples of the target species and spectra. These endmembers were used to apply the unmixing algorithm to all 6 weeks of imagery. See the Appendix for additional spreadsheets and graphics of data transformations.

Additional analysis took the form of multiband algorithms sensitive to spectral shape. The spectral shape equation [Eq. (1)] was adapted to be applied to the first derivative spectra to separate Aphanizomenon and Microcystis.⁷ The adapted equation is shown in Eq. (2) and is referred to as the derivative spectral shape (derivSS) equation. This was applied to two sections of the spectra, at 468 and 509 nm, where the shape of the derivative was observed to differ significantly from Aphanizomenon to Microcystis (Fig. 3 and Index 1)

Fig. 3

Derivative spectra of Aphanizomenon and Microcystis, with differing features at 468, 509, and 628 nm circled. This plot shows a subset of the full wavelength range of the hyperspectral microscope.

A third index was also employed, the slope of the derivative curve from 628 to 632 nm. This index was chosen because 628 nm was the location of the phycocyanin absorption feature for Aphanizomenon, so the index was positive, while for Microcystis, the phycocyanin feature occurred at a shorter wavelength and the index was less positive (data not shown). In addition, the slope of the derivative for green algae at this point is negative. These three indices served to demonstrate that the spectral features that appeared to differ between the two genera in region-of-interest (ROI) average spectra were consistent across all the imagery collected.

For some instances, a dataset was prepared consisting of single-cell ROIs from nearly all images collected in the final 6 weeks. Between three and six ROIs were designated in each image, resulting in a dataset with 165 samples of Aphanizomenon and 195 samples of Microcystis. In addition, ROIs for other organisms found in the samples, such as diatoms, green algae, and other cyanobacteria, were catalogued and included in this dataset. Average values from these ROIs were used to assess the ability of different algorithms to separate genera. The “ROI dataset” in the following discussion refers to this collection of spectra. For each week, a mosaic image was constructed with all the data collected at the highest magnification so classification methods could be tested on images with more spectral variation.

3.1.

Observations

Hyperspectral image cubes collected during the final 6 weeks of sampling, August 8th to September 15th, were used in analysis because a different data collection method was used for the first 4 weeks and the cubes are not comparable.

Gas vesicles are visible in both Microcystis and Aphanizomenon cells. Heterocytes (nitrogen-fixing cells) and akinetes (resting cells) were observed in Aphanizomenon filaments. Both Microcystis and Aphanizomenon were observed in week 1 (July 13, 2016) through week 7 (August 22), but week 8 (August 29) was dominated by Microcystis, and Aphanizomenon was less plentiful in weeks 9 and 10. This reduction corresponded well with recorded cell counts from these weeks (data not shown).

Various unspecified green algae were observed throughout Microcystis colonies, beginning the week of August 8 (week 5). An unidentified cyanobacterium was associated with Microcystis colonies in week 6. Protozoans were observed grazing on Microcystis during weeks 9 and 10, as the bloom died off. Pennate diatoms were also observed during these weeks. These observations potentially underestimate the diversity of the phytoplankton population because only a small fraction of the algae in the sample bottle was transferred to a slide.

In general, the reflectance spectrum for cyanobacteria consists of a Chl-a reflectance peak at 565 nm, phycocyanin absorption at 620 nm, and Chl-a absorption at 680. Both Aphanizomenon and Microcystis spectra include these features, and the spectra of green algae lack the phycocyanin feature. Figure 4 shows sample spectra from the major taxa present in the UKL samples.

Fig. 4

Example spectra from cells of Aphanizomenon (green), Microcystis (red), and green algae (blue). Reflectance at 565 nm and absorbance at 680 nm is indicative of chlorophyll a, while absorbance at 620 nm is indicative of phycocyanin.

Although the major spectral features were similar for all algae, there was significant variation with each genus. Aphanizomenon and Microcystis were both observed in a range of colors from yellow-green to blue-green.

It should also be noted that spectral areas of traditional vegetation analysis, the red edge and the near-infrared (NIR) slope, did not lend themselves to spectral distinction between the algal species. This was likely due to the mixed-media environment of partially submerged vegetation in water, and this could prove to be an important departure from standard hyperspectral vegetation analysis.

3.2.

Data Analysis

The linear spectral unmixing algorithm returned the most accurate results compared to other ENVI classification methods (Fig. 5), but still resulted in misclassified pixels. Figure 6(a) shows the average abundance values returned by the algorithm over single cell ROIs. Points above the $1:1$ line are classified as Aphanizomenon and points below are classified as Microcystis. Misclassified pixels could be dangerous from a false positive/false negative standpoint. Unclassified pixels are not useful, but in the hyperspectral microscope imagery, there were generally enough classified pixels in an area to infer the identity of the unclassified ones. This should be less of a problem with imagery of increasing areal extent, capturing more pixels.

Fig. 5

(a, d) Classification methods applied to normalized images, (b, e) PC transformations, and (c, f) spectral derivative transformations of the scene shown in Fig. 2(a). Both (a–c) the support vector machine classification and (d–e) the linear spectral unmixing algorithm were most successful when applied to derivative transformations. The unmixing maps have been optimized with contrast stretching. Scale bars are $100\mu \mathrm{m}$.

Fig. 6

(a) Average linear spectral unmixing values for the ROI dataset. (b) Average linear spectral unmixing values for the ROI dataset after derivative transformation. Points above the $1:1$ line are classified as Aphanizomenon and points below are classified as Microcystis. Red diamonds are cells identified in the imagery as Microcystis; green squares, Aphanizomenon; and blue circles, green algae.

Data transformations such as the spectral derivative accentuated the subtle, consistent differences between Aphanizomenon and Microcystis spectra. Figure 5(b) shows how the two cyanobacterial genera became more separable after the derivative transformation. However, the mapping algorithms still resulted in errors, and differentiated cells and color variations still led to some incorrectly classified pixels. For the ROI dataset, 70% of Aphanizomenon and 69% of Microcystis were classified correctly.

However, keep in mind that there are other cellular variations and structures within the matrix of features in the typical algal sample, not just the pure algal species. These components and variations have different structural components and different color variations and lead to some incorrectly classified pixels, especially when mapping algal species. Classification applied to PC transformations was less successful than the same process applied to derivative data, resulting in more misclassified pixels [Fig. 5(e)].

The derivative spectral shape indices were developed to determine if classification could be improved by focusing on key segments of the spectrum, and to determine if the unique spectral features observed for each genus were applicable across all the imagery collected. The indices focus only on the wavelengths at which Aphanizomenon and Microcystis differ the most, taking advantage of peaks and troughs that are accentuated by the derivative transformation. The derivSS(468) and derivSS(509) were observed to be more positive for Microcystis, while the second derivative at 628 nm is close to zero for Microcystis and positive for Aphanizomenon. In addition, the second derivative at 628 nm is markedly negative for green algae. The values of the derivative indices applied to the ROI dataset are shown in the scatterplot in Fig. 7. This scatterplot was used to develop thresholds for each class, which were applied to data cubes to construct classification images. Pixels for which all three indices match the criteria for a given class were assigned the designated color for that class, and pixels that failed a single criterion were left unclassified. Of all classification methods used, the method based on the spectral shape indices resulted in the fewest misclassified pixels. Figure 8 shows a mosaic of images from week 5, classified with linear spectral unmixing and with the derivative spectral shape indices. The derivative spectral shape indices resulted in fewer misclassifications than the linear spectral unmixing algorithm (Fig. 8). In the ROI dataset, 96% of Microcystis cells were classified correctly, and 100% of Aphanizomenon cells were classified correctly, in this case meaning that the number of pixels in the ROI classified as the observed cell genus was greater than the number of classified genus, disregarding unclassified pixels.

Fig. 7

Derivative spectral shape values of the ROI dataset. Class thresholds were established based on this plot and on a plot of the second derivative at 628 nm.

Fig. 8

Mosaics of week 5 imagery after different classification methods. Each image is labeled with the manually identified genus, AFA for Aphanizomenon and MC for Microcystis. (a) Linear spectral unmixing of normalized data. (b) Linear spectral unmixing of derivative data. (c) Linear spectral unmixing of PC data. (d) Derivative spectral shape index classification. Red pixels have been classified as Microcystis; green, Aphanizomenon; blue, green algae. All unmixing algorithms included the three genera and a background endmember. Contrast stretching was applied to optimize each unmixing map.