| L Tasiopoulos, M. Stefouli, Y. Voutos, Ph. Mylonas, E. Charou |
| Machine learning techniques in agricultural flood assessment and monitoring using EO and hydromorphological analysis |
| 13th EFITA International Conference, Digital Agriculture Web Conference, May 25-26, 2021 |
|
ABSTRACT
|
| Floods are among the most catastrophic natural disasters causing important and/or permanent damages to infrastructures and crops and livestock. Agricultural flood monitoring is important for food security and economic stability. In the future, climate change could exacerbate these phenomena by increasing the frequency of extreme and adverse meteorological events. Flood mapping could also serve other stakeholders and purposes such as risk management, land use and land management, emergency planning. Remote sensing (RS) is widely recognized as a unique source of data as it provides synoptic view over large areas. The constellation of Sentinel satellites is part of the Copernicus Earth Observation program led by the European Commission and operated by the European Space Agency. In this study, we propose a method to synergistically combine different types of data and processing techniques in order to achieve greater, more consistent and robust mapping accuracy compared to traditional approaches based on segmentation of single water/spectral index. The major long-term objective is to lay the foundation of an automated algorithm for mapping flooded areas requiring less a-priori sets and, above all, capable to cope with choices taken under imprecise information, compared to more traditional methods proposed in the literature. Data used are multi-temporal Sentinel2 data, orthophotos of 0.5m resolution, DEM of 5m resolution and ancillary land cover / use maps. Three sites were selected for algorithm set up and the assessment of mapping products that is areas affected by flooding during August / September 2020. The areas are located in Evia / Politika area, Cephalonia Pilareon municipality and Thessalia plain. The sites have been used for algorithm set up where training and testing pixels were extracted for i) definition of the “standing water”, ii) test of different multi-source soft integration operators and iii) validation of algorithm performance. Sites were selected to cover different conditions of standing water in order to capture variable spectral characteristics: flooded area is due to extreme heavy rainfall, river bed and flooded cotton fields. The application has as a result estimates of the total arable land that has been affected, that is, to find the percentage of land that is covered by the floods. In this work we developed an algorithm to automatic map flooded areas from multispectral S2 MSI images based on fuzzy set theory. Since the use of multisource data is recognized as the way to achieve improved global and regional water mapping, we propose an approach to integrate multiple spectral features. Rather than making an a-priori selection of the best (or a few) water indicator(s), we exploited redundancy provided by multiple spectral features. Machine learning is carried out with operators that can flexibly aggregate inputs of different nature: spectral indices, H/V, other types of data such as surface elevation gradient, single spectral bands, hydro morphologic parameters etc. The algorithm has been tested with input features taking into account the temporal dimension (pre – post event change detection) to further enhance flooded area mapping accuracy.
|
| 25 May , 2021 |
| L Tasiopoulos, M. Stefouli, Y. Voutos, Ph. Mylonas, E. Charou, "Machine learning techniques in agricultural flood assessment and monitoring using EO and hydromorphological analysis", 13th EFITA International Conference, Digital Agriculture Web Conference, May 25-26, 2021 |
[  PDF] [
BibTex] [
Print] [
 Back] |
