Plant-Based Contaminant Monitoring at SRNL: Analysis
The efforts of the collaborative plant-based contaminant monitoring project between Savannah River National Laboratory (SRNL) and the U.S. Army Corps of Engineers resulted in the development of a series of charts that analyze the data collected from the hyperspectral camera images.
The graphs below are an early proof of concept of the machine learning component of this project. A clustering algorithm known as “K-means” was applied across the entire spectrum of hyperspectral data provided by the camera, then a computer program was instructed to cluster “like” pixels within the image. Though these are the predominant representations, the clustering does not perfectly capture where each pixel resides in reality. The list below clarifies the area represented by each cluster:
- Cluster 1: The table upon which the pot of grass is sitting.
- Cluster 2: The grass within the pot.
- Cluster 3: Shadows from the grass cast on the table.
- Cluster 4: Additional material in the pot around the grass (e.g., dirt).
The color spectrum on the right of the image shows the separability of the identified clusters. Cluster 1 pixels reflect the most light at short wavelengths, followed by cluster 2, while clusters 3 and 4 reflect approximately the same amount of low wavelength light. Cluster 1 pixels show high reflectance up to about 350 nanometers while clusters 2, 3, and 4 show high correlation in wavelengths greater than ~250nm. The characteristics of each cluster’s reaction to light help to identify what contaminants are found within the grasses and at what concentration.
This work has the potential to make many contributions to the field of image processing in addition to this project’s primary goal of predicting contaminant uptake by the plants. Long-term applications of the project include other DOE sites, as well as commercial and industrial sites that could utilize a variety of different plant types that lend themselves to absorbing specific contaminants . Additional implementations of phytoremediation will improve clean-up efforts across the DOE complex through passive long-term monitoring that is both more affordable and accessible compared to other approaches.