TY - JOUR
T1 - Highly turbid and eutrophic small water bodies in West Africa well identified by a CNN U-Net algorithm
AU - de Fleury, Mathilde
AU - Grippa, Manuela
AU - Brandt, Martin
AU - Fensholt, Rasmus
AU - Reiner, Florian
AU - Kovacs, Gyula Maté
AU - Kergoat, Laurent
N1 - Publisher Copyright:
© 2024 The Authors
PY - 2025
Y1 - 2025
N2 - Although high-resolution multispectral optical imagery is increasingly being used to monitor continental surface waters more easily than ever before, there are still limitations to the methods used to extract water bodies. Detecting water becomes particularly difficult in the presence of aquatic vegetation or trees, or when spectral variations across the water surface are high. These limitations pose significant challenges in West Africa, where such cases are numerous, hindering the application of widely used methods and leading to a reduced quality of various existing datasets. As a result, the region lacks comprehensive information on the number of water bodies, their surface area, their spatial distribution and their typology. In this study, we propose a method based on a convolutional neural network based on a U-net architecture, which we apply to images from the Sentinel-2 multispectral instrument acquired in November 2020 and March 2018, corresponding to the maximum and minimum water area extent during the 2016–2020 period. We observe a much larger number of lakes than in current datasets, a large proportion of which are small and temporary. Overall, 29,265 water bodies were classified in November 2020 and 8,093 in March 2018 over an area of 1,340,450 km2 in the central Sahel, with sizes ranging from 0.002 km2 to 1,162 km2. In addition, a wide diversity of optical water types was found across the water bodies: hypereutrophic water bodies dominate, accounting for 67.9% in November 2020, followed by very turbid water bodies representing 26.1%. The Convolutional Neural Network U-Net algorithm successfully identified water bodies with aquatic vegetation or obscured by trees, as well as extremely turbid small lakes and reservoirs, which are often missing in global datasets. Such improved mapping capability has important implications for the monitoring of water resources and water quality, which are pivotal for the livelihoods of the region.
AB - Although high-resolution multispectral optical imagery is increasingly being used to monitor continental surface waters more easily than ever before, there are still limitations to the methods used to extract water bodies. Detecting water becomes particularly difficult in the presence of aquatic vegetation or trees, or when spectral variations across the water surface are high. These limitations pose significant challenges in West Africa, where such cases are numerous, hindering the application of widely used methods and leading to a reduced quality of various existing datasets. As a result, the region lacks comprehensive information on the number of water bodies, their surface area, their spatial distribution and their typology. In this study, we propose a method based on a convolutional neural network based on a U-net architecture, which we apply to images from the Sentinel-2 multispectral instrument acquired in November 2020 and March 2018, corresponding to the maximum and minimum water area extent during the 2016–2020 period. We observe a much larger number of lakes than in current datasets, a large proportion of which are small and temporary. Overall, 29,265 water bodies were classified in November 2020 and 8,093 in March 2018 over an area of 1,340,450 km2 in the central Sahel, with sizes ranging from 0.002 km2 to 1,162 km2. In addition, a wide diversity of optical water types was found across the water bodies: hypereutrophic water bodies dominate, accounting for 67.9% in November 2020, followed by very turbid water bodies representing 26.1%. The Convolutional Neural Network U-Net algorithm successfully identified water bodies with aquatic vegetation or obscured by trees, as well as extremely turbid small lakes and reservoirs, which are often missing in global datasets. Such improved mapping capability has important implications for the monitoring of water resources and water quality, which are pivotal for the livelihoods of the region.
KW - CNN
KW - Typology
KW - Water bodies
KW - Water extraction
KW - West Africa
U2 - 10.1016/j.rsase.2024.101412
DO - 10.1016/j.rsase.2024.101412
M3 - Journal article
AN - SCOPUS:85213005663
SN - 2352-9385
VL - 37
JO - Remote Sensing Applications: Society and Environment
JF - Remote Sensing Applications: Society and Environment
M1 - 101412
ER -