Search the EPA Research Database

Project Search Result

Project Code [GOIPG/2020/790]

This information is correct as of today and is updated from time to time by the EPA to reflect changes in the management of the project. Please check back regularly for updates.

Project title

A multi-data machine learning approach to identifying, mapping and characterising sinkhole populations in karst environments

Primary Funding Agency

Irish Research Council

Co-Funding Organisation(s)

Environmental Protection Agency

Lead Organisation

University College Dublin (UCD)

Lead Applicant

n/a

Project Abstract

Karst environments form by dissolution of soluble rocks, and they supply drinking water to about 25% of the world�s population. Landscapes in karst environments are characterised by enclosed depressions of 1-500 m diameter called sinkholes (or dolines). Since they commonly form by subsidence, sinkhole collapse is a serious geohazard in karst environments. Additionally, sinkholes often function as entry points for contaminants into karst groundwater supplies. For assessing these hazards, accurate recording of sinkhole distribution and attributes at the metre-scale is essential. The ever-increasing quantity and accessibility of high-resolution Earth surface data from drones, aircraft and satellites provides an opportunity for improving sinkhole hazard assessment in karst environments. However, correspondingly large data volumes hinder the generation of reliable sinkhole databases in a timely fashion. Manual sinkhole identification is highly labour-intensive, but is presently the default method because automated mapping methods are not currently of comparable accuracy. This is in large part because present automated methods do not exploit combinations of various Earth surface datasets. This project aims to develop a new automated method of identifying sinkholes by using machine learning techniques (Neural Networks) to analyse multiple datasets, each recording different sinkhole characteristics, simultaneously. The machine learning approach will be trained, validated and applied to datasets comprising optical satellite imagery and 3D topographic models of sub-metre resolution from a range of karst environments. We will study the derived sinkhole distribution and attributes to better understand links between different morphologies and formation processes in these environments. Ultimately, we aim to use the new method to create an open-source tool to enable robust and rapid mapping of karst depressions within Geographic Information Systems. In the long run, this tool would aid assessment of groundwater vulnerability and planning of infrastructure projects as part of managing and sustaining natural resources in karst environments.

Grant Approved

�96,000.00

Research Hub

n/a

Research Theme

Not relevant

Start Date

01/01/2021

Initial Projected Completion Date

31/12/2024