Deep learning waterways for rural infrastructure development

TL;DR

A computer vision model is built to learn the location of waterways in the United States, based on high resolution satellite imagery and digital elevation models, and then deployed in novel environments in the African continent, showing promise for capturing humanitarian needs and planning for social development in places where cartographic efforts have so far failed to deliver.

Abstract

Surprisingly a number of Earth's waterways remain unmapped, with a significant number in low and middle income countries. Here we build a computer vision model (WaterNet) to learn the location of waterways in the United States, based on high resolution satellite imagery and digital elevation models, and then deploy this in novel environments in the African continent. Our outputs provide detail of waterways structures hereto unmapped. When assessed against community needs requests for rural bridge building related to access to schools, health care facilities and agricultural markets, we find these newly generated waterways capture on average 93% (country range: 88-96%) of these requests whereas Open Street Map, and the state of the art data from TDX-Hydro, capture only 36% (5-72%) and 62% (37%-85%), respectively. Because these new machine learning enabled maps are built on public and operational data acquisition this approach offers promise for capturing humanitarian needs and planning for social development in places where cartographic efforts have so far failed to deliver. The improved performance in identifying community needs missed by existing data suggests significant value for rural infrastructure development and better targeting of development interventions.

Figures (6)

• Figure 1: Potential data gaps in waterways representation in select countries in Africa. a-c) show the countries (Liberia, Côte d’Ivoire, Ethiopia, Rwanda, Uganda, Tanzania, Zambia, Kenya, Spain, France, Belgium, Netherlands, Germany, Switzerland, Austria, Italy) or watersheds used for comparisons. d) shows the correspondence between WaterNet and existing widely used, and gold standard datasets, showing substantially weaker coverage in African geographies, than countries in western Europe or watersheds in the USA, as well as lower coverage on lower stream orders. The percentage of WaterNet inner segment points at the specified stream order (O1, O2, O3, all) within 0.002 degrees of the corresponding dataset are shown.
• Figure 1: Illustrative examples of missing waterways in NHD data. The bounding boxes for a, b, and c are (-87.8104, 45.1841, -87.7965, 45.1955), (-94.2725, 46.3569, -94.2522, 46.3680), and (-88.5530, 44.6436, -88.5177, 44.6550) respectively. None of the water in these bounding boxes were labeled in the NHDPlus data.
• Figure 2: Example model outputs from the deep learning drawn waterways. a, b, and c are the model’s raw raster outputs with the following respective bounding boxes, Ethiopia (38.94, 11.9191, 39.46, 12.1791), Rwanda (30.04, -1.98, 30.56, -1.72) and Côte d’Ivoire (-5.4, 7.5517, -4.88, 7.8117). c, d, and f are the deep learning model’s vectorized outputs (in blue) with TDX-hydro (in orange).
• Figure 2: Fine tuning model for humanitarian use-case. Major 2019 South Sudan flooding leading to human displacement. Identification of event with (left) and baseline (right) adjusting fcode weights for swamps.
• Figure 3: Community requests for infrastructure are well captured by our model. Comparisons with OpenStreetMap (OSM) and TDX Streams are given for comparison, showing the gap in community needs and what current scientific mapping exercises have been able to map.