Conference Program

Ensemble Learning Models for Biomass Estimation and Species Classification of Intertidal Macroalgae Using In-situ and Remote Sensing Spectrometry (310249)

Stefan Claesson, Nearview, LLC 
Rosemary Danaher, University of New Hampshire 
Matthew Duckett, University of New Hampshire 
*Ernst Linder, University of New Hampshire 
Khem Veasna, University of New Hampshire 

Keywords: Machine Learning, Coastal Resources, Images, Segmentation, Drones, Satellites.

Macroalgae in the intertidal zone of the North Atlantic have become an important resource for human consumption, animal feed, fertilizer, and other uses. Commercial harvesters, marine scientists and resource managers have an intense interest in assessing the state of this resource and to ensure sustainable practices. Traditional "field-based" in-situ methods for estimating biomass and delineating the main species, such as ascophyllum and fucus, are time-consuming, expensive, and imprecise due to sparseness. We present an integrative approach to improve on sparse field-based estimates by augmenting them with lab and in-situ generated high resolution spectrometry and remotely sensed low resolution spectral reflectance at the 1 m^2 scale from unoccupied Aerial Systems (i.e. drones) and at the regional scale by using (10^2 m) satellite information such as from Sentinel-2. Our research will be incorporated in an interactive tool where users can obtain predictions for a particular coastal area. Our statistical and ML modeling examines translations across platforms and resolutions. While controlled lab experiments provide species classification and biomass estimation with high precision and accuracy, the noisier field environments degrade the quality of similar estimates. With extensive testing of different learning models (discriminant analysis, support vector machine, neural nets, random forest), we found that combining these into an "ensemble model" provides the best predictions of the noisier field ground-truth. Expert-based drone-derived image segmentation by species classification in several coastal sites in Maine (USA) have been successfully validated by our ensemble models. We also predict biomass (weight) by species using "depth" predictions via analogous ensemble models, and a lab-derived calibration curve of the depth-weight relationship. We also examine up-scaling to the regional scale for assessment, based on lower resolution satellite reflectance data.