Semi-supervised LULC segmentation on BD-Sat using Cross Pseudo Supervision

Abstract

Semantic segmentation in Land Use Land Cover (LULC) tasks is getting increasingly relevant. This paper uses a semi-supervised semantic segmentation learning model on a map dataset of parts of Dhaka Division and its corresponding annotations collected from BD-Sat [1]. The semi-supervised semantic segmentation model that is used in this project is called cross pseudo supervision (CPS) [2]. Two identical neural networks are used with different initializations and are used with both the labeled and the unlabeled data. A semi-supervised approach is used in this project because we intend to make use of the additional unannotated map images to further enhance the effectiveness of semantic segmentation. This paper also compares the results of the semi-supervised model with that of the supervised model obtained from the same dataset. It thereby shows that cross pseudo supervision works remarkably well when compared to its supervised alternatives.