Onchocerciasis Vector Habitat Detection Using Remote Sensing and Machine Learning Methods

The strategy to eliminate onchocerciasis, a parasitic neglected tropical disease transmitted through bites by blackflies of the genus Simulium, has traditionally relied upon multiple tactics that all, to some degree, require geospatial information to accurately map populated areas and vector monitoring and/or control. Statistical “search space reduction” techniques can be used to predict the location of likely vector habitat, thereby increasing program efficiency by optimizing limited resources such as travel costs, labor, and time. Developing an effective elimination endgame also relies upon accurately mapping endemic foci and vector habitat suitability. This study explores if traditional and geostatistical methods, in conjunction with computer vision and machine learning analysis of remote sensing data, can optimize the detection of the Simulium vector habitat. Specifically, where and when is habitat suitable to Simulium breeding? In partnership The Carter Center (TCC), we used river prospection data provided by TCC’s Ethiopia Onchocerciasis Elimination Programme (OEP) that identify the date, geographic coordinates, and the presence or absence of Simulium aquatic forms/breeding habitat (the outcome/dependent variable). We calculated variables to describe those geographies at the prospection date using open-source atmospheric data from the Japan Aerospace Exploration Agency (JAXA) and the UC Santa Barbara Climate Hazards Center, and imagery from Landsat and Sentinel. Variables are calculated within a 14-day temporal buffer and in a 270m geographic buffer. The team also calculated all dynamic features at 2, 4, 6, 8, 26, and 52-weeks before prospection to capture seasonality. We developed two Histogram Gradient Boosting Classifier models. One model predicts two weeks in advance of prospection using data from a single time and the second model predicts 6 weeks in advance of prospection using multiple historical data points. Models exhibit an overall accuracy of 75/79%, sensitivity of 74/73%, and specificity of 76/83%, respectively, on the testing dataset. A flexible model-agnostic workflow generates future prospection maps. The two models have different variable importance, with the 2-week model aligning more closely to existing literature but the 6-week model possibly being more operationally useful by the OEP. We overcame technical challenges to build a replicable and robust process, including analyzing large amounts of imagery over a ten-year period and addressing missing values resulting from cloud cover. Both models and their assumptions need to be validated, tested, and refined under field conditions in Ethiopia, in conjunction with the OEP and TCC team. We present a proof-of-concept illustrating the feasibility of using remote sensing data and machine learning to predict the presence of Simulium aquatic forms/breeding habitat. Planned next steps include (1) field-testing and in situ validation of the model(s), (2) model refinement, (3) modularization of the process into a deployable system, and (4) evaluation of climate change on habitat evolution.