Combining night time lights in prediction of poverty incidence at the county level

Abstract

Long-term poverty data can support accurate decision-making. This study demonstrates an accurate and reliable method for identifying poverty areas and predicting poverty incidence based on night time light remote-sensing data and machine learning methods. Using data of poverty counties and poverty incidence in Guizhou Province of China as the training dataset, we show how to use machine learning to identify poverty counties and predict poverty incidence in the Yunnan-Guangxi-Guizhou Rocky desertification area. The identification accuracy of poverty-stricken counties was 76.5%. The root mean squared error, mean absolute error, and R2 values of the poverty incidence rates were 5.01, 4.04, and 0.60, respectively. Using data from 2015 to verify the trained model, the R2 value of the predicted and actual values of poverty incidence reached 0.95. With the progress in machine learning and night light remote sensing, poverty mapping combined with night time lights and machine learning can compensate for the data gap in deprived areas and provide a decision-making basis for sustainable development in poverty-stricken areas.