Poverty Mapping Using Satellite Imagery

Abstract

The absence of trustworthy data in poor nations is a significant barrier to disaster assistance, food security, and sustainable development. For instance, information on poverty is often hard to get by, poorly covered, and labor intensive to gather.As opposed to that, data from remote sensing, like high resolution satellite photography, is turning more widely available and rea sonably priced. Unfortunately, because of how poorly formatted this data is, there are currently no methods for automatically extracting insightful infor mation that may be used to guide humanitarian activities and advise policy. From highly detailed satellite data, large-scale socioeconomic factors will be extracted., we suggest a unique machine learning approach. The key difficulty is the lack of high-quality training data, which Pakistan lacked, making it challenging to use cutting-edge methods like convolutional neural networks (CNN). Therefore, we suggest a transfer learning strategy that uses nighttime light intensities as a rich surrogate for data. In order to forecast nighttime lights from daytime footage, While concurrently learning characteristics that are useful for predicting poverty, we train a fully convolu tional CNN model.. With the exception of nighttime lights, the model learns filters to distinguish between various terrains and man-made features, such as highways, buildings, and farmlands. We show that these learned traits are quite useful for mapping poverty and even come close to matching the prediction capabilities of field survey data.