Convolutional neural network with U-Net architecture to detect buildings in satellite imagery for statistical purposes

Abstract

Statistics Indonesia (BPS) carries out a population census every ten years, which provides data on the number, composition, distribution, and characteristics of the Indonesian population. However, the vast geographical area and the variety of tribal communities pose a challenge in adequately conducting this activity. Over the years, BPS has taken various ways to minimize under coverage issues, such as digitizing, tagging, and categorizing each region into local environmental units (SLS) or non-SLS. The population census is carried out in areas categorized as SLS because non-SLS are uninhabited. However, in reality, there are conditions where the non-SLS areas, such as inland and forest, are inhibited by tribal people, thereby making it difficult for enumerators to collect data. Therefore, this study proposes an innovative approach to underpin the population census in non-SLS areas using a building detection approach. It is assumed that the presence of a building in an area indicates the possible presence of people living there. The research developed nine CNN models with U-Net architecture capable of performing semantic segmentation to detect buildings on Sentinel 2 and Landsat 8 satellite imagery using some bands. The result showed that the U-Net architecture is the best model with input dimensions of N $\times$ 64 $\times$ 64 $\times$ 10, where N is the amount of data capable of detecting buildings in Sentinel 2 satellite images. Furthermore, the model produces an F1 score on the validation data of 0.7282 and has been trained for 50 epochs (repetition) with a dashboard developed to visualize the detection results with a plugged-in model.