Paper details

Abstract

To accurately identify the areas most vulnerable to floods, physical and social vulnerability information of the population at risk is required. However, in developing countries, this information is often unreliable, unavailable or inaccessible. Geospatial data such as remote sensed imagery and household surveys can be collected to fill this data gap. This data coupled with artificial intelligence (AI) enables interventions or response at a household level. However, unintended biases have been observed in AI algorithms used for image recognition, credit scoring, and criminal recidivism. The use of AI should not amplify the biases present in the baseline situation where no geospatial data and AI is used. As a case study, we analyze a geo-intelligence workflow that integrates AI, remote sensed imagery, mapillary street view images, and household survey data to characterize housing stock vulnerability to flooding in Karonga district, Malawi. Our research demonstrates possible biases in the geo-intelligence workflow. For example, in data collection and preparation (representation, measurement, and aggregation bias), model inference (evaluation and hyper-parameter bias) and deployment (deployment bias). The identification of these biases can steer subsequent data collection efforts and improvements to the workflow as well as create awareness and in some cases help mitigate biases.