The effects of climate change on water resources have led to increased water-stress in rural societies in sub-Sahara Africa (SSA), with adverse effects on humans, livestock and crop production. With low resource availability and long distances to permanent water sources, these societies are reliant on seasonal shallow ponds to satisfy both household and livestock water needs.
Seasonal availability, coupled with high siltation rates, further increases the uncertainty in water availability in SSA. Earth observation (EO) with technological advancement in feature detection through artificial intelligence (machine learning) provides the best way to map and monitor changing water resources.
However, cloud contamination and atmospheric interference limit the use of optical EO in water resources monitoring due to inconsistency in data availability, especially in the tropics. Active microwave sensing data primarily synthetic aperture radar (SAR) with its all day-and-night data acquisition solves the problem of data availability for continuous water resources monitoring.


Using SAR data with deep learning, to detect and map the location and availability of shallow water ponds to local communities and monitor their status.
This information, integrated into apps used by tourists, can identify watering points for wildlife; for health organizations, to identify breeding grounds for mosquitoes; etc. Additionally, for the preparation of food-security projects; prevention of human-wildlife conflicts caused by limited water resources or the creation of early warning systems for water stress in local societies.



PyroSAR framework was used; this is a python framework for large-scale SAR satellite data processing. This framework was used to process the Level 1c GRD (ground range detected) C-band Sentinel -1A & B SAR datasets, with IW (interferometric wide swath) acquisition mode.

Training samples were generated from the Western parts of Kenya using ArcGIS. The images are then divided into small chips of 256 x 256 pixels with a 128 x 128 overlap, for easy usage in the network during the training of the models.
The images are then exported to deep learning with KITTI Rectangles from ArcGIS.

Data Division

Datasets used were divided into 3, given the bi-modal nature of rainfall received in W. Kenya. (1) Wet season — short rains — January to April. (2) Dry season- dry- May to August (3) Wet season — Short rains- September to December.

Deep Learning with Neural Networks

Neural The relationship between SAR backscatter and water bodies were explored in Neural Networks was implemented within Google’s Tensorflow library in Python to extract water bodies in Western Kenya, then Normalised. The training was done on 2623389 samples, using Neural Networks was implemented within Google’s Tensorflow library in Python.


Only water areas were extracted and the rest of the features were classified as one feature. As shown in the image below, blue areas marked all the water points detected. These areas, through time series analysis, are compared to determine the progressive changes in the water area.

Detected water ponds (points)

Tech company based in Berlin with interest in the African Market