Identifying upland blanket bog drainage channels using machine learning and very high resolution aerial remote sensing.

Peatlands only cover 3-4% of Earth’s terrestrial surface yet are globally important carbon stores, hydrological regulators and biodiversity hotspots. Blanket bog, a type of peatland, is a rare ecosystem type within the EU. Active and degraded blanket bogs are spatially abundant in Ireland covering ~11% of the country. EU Natura 2000 protected areas includes ~167,000 ha of active blanket bog, of which 93% is in Ireland. Yet despite peat soils extending to 1.66 Mha of Ireland (~23.3%), more than 90% of these peatlands have experienced anthropogenically induced degradation. Land use change and digging artificial drainage challenges are key mechanisms of degradation: impacting biodiversity, hydrology and resulting in carbon and greenhouse gas emissions. Consequently, few actively functioning Irish upland blanket bogs remain. Rewetting, rehabilitation and restoration could be conducted to improve local biodiversity, water regulation and water purification, and reduce emissions.
Peatland rehabilitation relies on effective water table management, including blocking drainage ditches, to raise the local water table to promote  wetland vegetation. Drainage ditches are extensive in Ireland and extend to thousands of kilometres, but manual mapping is expensive and time consuming. Still, these drainage ditches must be mapped to identify potential areas for rehabilitation. Potential areas to be restored must be mapped  and quantified under the EU Nature Restoration Law and Biodiversity strategy. 
We aim to: (1) adapt a methodological workflow using deep learning and very high resolution aerial imagery to map artificial drainage ditches in Irish upland blanket bogs ; and (2) combine the model with other GIS analyses to indicate upland bog rehabilitation potential. Our work adapts previous raised-bog drainage mapping models, and utilises recent peatland and peaty soil extent maps to delimit the analysis to these regions. 
The model was tested in the Wicklow Mountain upland blanket bog region. Initial results show that the model is effective at recognising linear drainage ditches on upland blanket bogs regardless of depth. Early analysis depict 464km of drainage channels in Co. Wicklow (East Ireland). Incomplete or interrupted drainage ditches may indicate that some drains are over-grown or have filled in with peat. . Current Completeness, Correctness and Quality (CCQ) accuracy assessment  indicates that the model identified drainage ditches with ~73% Completeness, ~94% Correctness and ~70% Quality. False positives seem to be limited to deer tracks or gullies. The same approach will be implemented Co. Mayo and Co. Sligo (West Ireland) demonstrating the transferability of these methods and potential for upscaling to a national level. The outputs from this study will inform policy, governance and practice as these bodies work towards meeting peatland restoration targets indicated in they implement EU and National Law.