Abstract:
In order to make proper policies, measurement of poverty needs to be accurate, but conventional survey-based instruments like the Demographic and Health Survey (DHS) are not only expensive, but also scarce and sparse spatial-resolution instruments. This paper will look at the question of whether the geospatial features of poverty in Bangladesh can be estimated using freely available features as an alternative. Based on the ground truth (DHS 2018 household recode data), several geospatial indicators, including Night-Time Lights (NTL) and Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), rainfall, land cover distribution, and Point-of- Interest (POI) density, were obtained including 672 survey clusters on Google Earth Engine and other open sources. Two measures of poverty were tubed, such as the Make-go of Wealth Index and Wealth Quintile, which were predicted by a series of machine learning models, such as the Generalized Least Squares (GLS), Random Forest Regressor (RFR), Support Vector regression (SVR), the XGBoost, the KNN, Decision Tree, Gradient Boosting, and Multi-layer perceptron (MLP) models. The R2, RMSE, MAE evaluation of the Wealth Index and Accuracy, Precision, and Recall evaluation of the Wealth Quintile prove that geo deterministic based models can be used to have a significant approximation of the official estimates of DHS. Random forest performed better among all the other models with the lowest RMSE of 365.220439 and the highest R2 of 0.759106. Compared to the other predictive models which had an accuracy of 69 and 53 percent, the predictive model Random forest was the most appropriate to be used in mapping poverty with a 86 percent accuracy. The feature importance analysis revealed that NTL, land cover, NDVI and POI accessibility have the highest role in predicting poverty.
