Self-learning Algorithm for In-Situ Identification of Toxic Algae
To make our algae sensor suitable for long-term monitoring in coastal areas, I developed a self-learning algorithm that analyses the current relative abundance of algal classes in real time and predicts algal blooms, especially those of potentially toxic algal classes.
Goal
Often it is not the problem to acquire data – but finding the hidden information in a larger set of data. Asking the right questions and analyzing the data in detail will give you relevant insights. The goal is finding the right questions and answer them so you can reliably build up on them.
Result
By using data mining, statistics, and machine learning techniques, I find patterns even in heterogeneous and complex data sets. Based on this information, I write an algorithm, probability estimation, or statistical model that helps you understand your situation and possible options to choose the best one for you.
GitHub
By clicking the icon, you will be redirected to GitHub, where you can download the software for Windows and macOS.
ALGAE – A DIVERSE WORLD
The term algae encompass a heterogeneous and diverse collection of different phototrophic organisms widely spread all over the world occurring both in terrestrial environments and in water bodies. The individual algae species are extremely diverse in their morphology and range in size from few micrometers to several meters long. The most common definition of algae is given by Lee in 1980:
“The algae are thallophytes (plants lacking roots, stems, and leaves) that have chlorophyll-a as their primary photosynthetic pigment and lack a sterile covering of cells around the reproductive cells.”
WHY MONITORING ALGAE?
Algae are at the very bottom of the food web and present therefore an essential source of organic matter and oxygen necessary for the metabolism of other organisms. In addition, algae possess the capability to respond rapidly and predictably to a range of pollutants, making them an interesting and sensitive indicator of changing environmental conditions.
However, at the same time, some algal species themselves can pose a severe threat to their surroundings, when they increase in biomass forming so-called algal blooms. In this state, algae can cause unfavorable conditions for marine life and humans, causing adverse effects on public health or on economic and ecological structures. Potential risks caused by algal blooms are:
-
clogged gills or lungs of fish and marine mammals
-
zones of low oxygen or even hypoxic dead zones in the water column
-
diseases such as shellfish poisoning syndromes or even death from biotoxins when ingested by fish, marine mammals, and humans
Over the years, the evident increase in intensity, frequency, and geographic distribution of algal blooms widely spread all over the world has focused attention on the intrinsic physiological, ecological, and toxicological aspects of individual algae species involved. Thus, it is worth studying the occurrence of algae, especially in coastal environments as well as their bloom dynamics including their bloom composition and related biological processes at a temporal and spatial scale.
A SPECTRAL APPROACH TO MONITORING ALGAE
As mentioned earlier, algal species can vary widely in morphology and size, making it extremely difficult to determine the relative abundance of algal groups in real-time with simple and low-cost instruments and to monitor them in-situ.
Pigments are substances produced by living organisms that are colored. A photosynthetic pigment or accessory pigment is a pigment present in chloroplasts or photosynthetic bacteria that absorbs the light energy necessary for photosynthesis. Examples of photosynthetic pigments are carotene, xanthophyll or chlorophyll-a/b.
However, when a chemist looks at algae, we see not (only) morphology and size, but also pigments. Do you remember? As phototrophs, algae perform photosynthesis to maintain their energy household – quite similar to plants. In all the different processes occurring during photosynthesis, pigments play a central role in converting sunlight into chemical energy that can be used by algae. Due to their central role, researchers have studied these pigments extensively and have assigned them a specific theory – known as pigment-based chemotaxonomy. This theory assumes that certain photosynthetic pigments are restricted to only one or a few groups of algae (taxa). These pigments are called marker or diagnostic pigments and mainly characterize the overall excitation spectrum. Therefore, marker pigments are well-suited for identifying and quantifying the composition of algal blooms.
The absorption spectrum of various marker pigments (top) and the normalized excitation spectrum of a diatom and a cyanobacterium (bottom). Left: The absorption spectrum of marker pigments is compared with the excitation spectrum of a diatom to illustrate that marker pigments significantly affect the overall excitation spectrum of algae. Right: Comparison of the excitation spectrum of diatom with that of cyanobacteria.
Based on these insights, we have equipped our algae sensor with 8 different LEDs that target the excitation of relevant marker pigments to reliably identify relevant algae groups. Read more about our low-cost miniature algae sensor here.
MULTIVARIATE PATTERN RECOGNITION FOR ALGAE CLASSIFICATION
Our algae sensor equipped with 8 LEDs records the fluorescence intensity of the algae sample upon excitation, meaning that each algae species is described by these 8 data points:
For in-situ identification, we required a self-learning algorithm with the lowest possible computational power that works on small single-board computers. Additionally, thanks to our cooperation partners, we were able to record algae species from different groups and could therefore build up training data. These training data were then used to screen for the best possible (supervised) algorithm.
Fisher's linear discriminative algorithm (LDA) was found to be the algorithm with the best performance parameters. LDA is a supervised pattern recognition algorithm that searches for a linear combination of measurement features to best separate the different groups. As a criterion for group separation, the variance between groups ( ) is compared with the variance within a group ( ) in the following way:
with is the vector of discriminants representing a weighted linear combination of measurement features , and are the within- and between-scatter or covariance matrices of the data matrix.
Upon determining a linear combination based on the given training data, the same linear combination is applied to the unknown sample, which is then re-localised relative to the training data. To assign the sample to a group, the (Mahalanobis) distance between the sample and the center of each group is computed and expressed as the probability of group membership using Gaussian distribution.
Project Gallery
A small slideshow to illustrate the pipeline for algae classification. The entire algorithm is combined in a software.
Do you want to hear more about data analysis? Please get in touch with me here.
References
Relevant articles and books
-
ZiegerSE, Seoane S, Laza-Martínez A et al. Environ. Sci. Technol. 2018, 52(24), 14266–14274
-
Zieger SE, Multiparametric Sensing – from Development to Application in Marine Science (Dissertation, 2019)
-
McCormick, P. V.; Cairns, J., J. Appl. Phycol. 1994, 6 (5–6), 509–526
-
Hallegraeff, G. M. A, Phycologia 1993, 32 (2), 79–99
-
Pitois, S.; Jackson, M. H.; Wood, B. J. B. Int. J. Environ.Health Res. 2000, 10(3), 203–218
-
Waeber, M.-L. T.; Bakker, E.; Nardin, et al. FP7-OCEAN-2013 - SCHeMA: Integrated in Situ Chemical MApping Probes. In OCEANS 2015 - Genova; IEEE: Genova, 2015; pp 1–5
-
Anderson, D. M.; Andersen, P.; Bricelj, M. V.; Cullen, J. J.; Rensel, J. J. E. APEC Report # 201-MR-01.1 59; Asia Pacific Economic Program and Intergovernmental Oceanographic Commission of UNESCO: France, 2001; p 268
-
Phytoplankton Pigments: Characterization, Chemotaxonomy and Applications in Oceanography; Roy, S., Llewellyn, C., Egeland, E. S., Johnsen, G., Eds.; Cambridge University Press: Cambridge, 2011
Webpages