Detection of roof type in rural Tanzania using high-resolution satellite imagery and convolutional neural networks

Abstract

The United Nations (UN) efforts to completely eliminate poverty by 2030 have been less successful in rural areas of Sub Saharan Africa compared to other parts of the world. This is partially due to infrequent and incomplete collection of census and other poverty-predictive data necessary to monitor progress and target resources to the communities most affected by poverty. The recent increase in available satellite data and development of machine learning (ML) tools could provide a way to increase accuracy and frequency of data collection for the improved monitoring of socioeconomic change. This study aims to use Convolutional Neural Networks (CNNs), a type of ML, to automatically extract roof type information from high-resolution satellite imagery of Mbola, Tanzania as changes in roof material are thought to be a spatially-predictive indicator of socioeconomic change. While the model reaches an accuracy of 85% for the detection of metal roofs in Mbola, this accuracy drops when the model is applied to different countries and time-periods, and thatch roofs are significantly more difficult to detect. This study highlights the need for further development of ML analysis techniques to better understand patterns extracted from data and to improve the efficiency of the training process. The development of more ML training data specific to African Earth Observation (EO) tasks is key before CNNs become a consistently viable tool for the monitoring of poverty in African countries.