Authors:
Ioannis Manakos
1
;
Malak Kanj
2
;
Michail Sismanis
1
;
Ioannis Tsolaikidis
3
and
Chariton Kalaitzidis
2
Affiliations:
1
Information Technologies Institute, Centre for Research and Technology Hellas, Thessaloniki, Greece
;
2
Department of Geoinformation in Environmental Management, Mediterranean Agronomic Institute of Chania, Chania, Greece
;
3
Lake Kerkini Management Authority, Kerkini, Serres, Greece
Keyword(s):
Inundation Mapping, Automatic Thresholding, Sentinel-2, Sentinel-1, Kerkini Lake, Wetland.
Abstract:
Satellite data may support management of wetland areas for monitoring of the inundation seasonality. Previously successful in Doñana and Camargue Biosphere Reserves, this study examines the transferability of unsupervised inundation mapping through automatic local thresholding in discriminating inundated areas from non-inundated ones in Kerkini Lake. Nine different alternatives of this approach are employed on Sentinel-2 (S2) Level-2A images (2016-2019). The best fit alternative was derived by the validation against local and on-site registered attributes. To overcome unfavourable atmospheric conditions, Sentinel-1 (S1) images were examined in tandem with derived S2 inundation maps (S2m), using the best fit alternative. Two S2m, one preceding and one following a target S1 image, were used to train random forest models (per pixel) to be applied to the target S1 image and derive the respective inundation map (S1m). S1m was validated against a S2m for the same date; not previously used
in the training process. Classification performance reached k [0.77-0.94] and overall accuracy [88.05-97.16%] for the S2m. The evaluation of S1m showed k of 0.99 and overall accuracy between 99.71-99.88%. Automation of the process and minimum human interference supports its usage by non-specialists, e.g. for Protected Areas management.
(More)