A Satellite-derived Peatland Ecotype Classification Method Using Artificial Neural Network Hierarchical Ensembles

Being able to identify, map and monitor areas of different ecological quality of peatland habitats, or ecotypes, provides important information on spatial peatland condition, the potential for restoration of degraded areas and ecotype carbon (C) emission and/or sequestration capacity when coupled with known C-flux factors. Regular and accurate mapping of such ecotypes is also a requirement under the European Union (EU) Habitats Directive, and will be required in some form to help guide the framework and implementation of the upcoming EU Nature Restoration Law.

Although the most precise way to identify the presence of certain ecotypes is via in-situ surveying, this approach clearly suffers from scaling issues, as it is only feasible in small selected peatlands (or even portions of them) and requires a lot of resources, e.g. skilled domain experts and time. A solution might come from remote sensing and Earth Observation technologies, which have been increasingly utilised to map the occurrence and extent of peatland environments in recent years. With this respect, the European Space Agency's Copernicus Program's Sentinel-2 satellite constellation could be a viable data source, as it allows for a multi-spectral, systematic and regular coverage of land surfaces with a spatial resolution up to 10 square metres and 5 days of revisit frequency. Nevertheless, the remote detection and mapping of ecotypes within the peatland complex itself is relatively under-studied and there is no currently accepted method that can be deployed at landscape scale. 

In this work we present a rather simple machine learning pipeline for ecotype detection at scale. The focus of this study are lowland peatlands, or raised bogs, in the Republic of Ireland. The pipeline assumes the existence of ground truth ecotype data (for machine learning training purposes), raised bogs map boundaries (shapefiles) and Sentinel-2 imagery. Both training, testing and validation datasets undergo the same pre-processing procedure. In the training step we train an ensemble of binary classifiers - specifically one multilayer perceptron network per ecotype - organised in a hierarchical fashion, to reduce the complexity of the problem. The ecotype classification would be done in a cascade - in accordance with the hierarchy - via canonical ensemble learning classification. 

The preliminary results gathered not only seems to indicate that our approach could provide reliable estimations about raised bog ecotype composition at scale, they also highlight the potential need for seasonal ensembles. Furthermore, we will present the results of a crowdsourcing experiment, in which domain experts were: (1) presented with ecotype map images, resulting from the inference of a plethora of ensemble classifiers of different settings and hyperparameters, and (2) asked to cast a vote on which image most closely resembled the related ground truth image.