Effect of Label Noise on Multi-Class Semantic Segmentation: A Case Study on Bangladesh Marine Region

Abstract

The volume and availability of satellite image data has greatly increased over the past few years. But, during the transmission and acquisition of these digital images, noise becomes a prevailing term. When preprocessing the data for computer vision tasks, human experts often produce noise in the labels which can downturn the performance of learning algorithms drastically. This study is directed toward finding the effect of label noise in the performance of a semantic segmentation model, namely U-net. We collected satellite images of the Bangladesh marine region for four different time frames, created patches and segmented the sediment load into five different classes. The U-Net model trained with Dec-2019 dataset yielded the best performance and we tested this model under three types of label noise (NCAR – noise completely at random, NAR – noise at random and NNAR – noise not at random) while varying their intensity gradually from low to high. The performance of the model decreased slightly as the percentage of NCAR noise is increased. NAR is found to be defiant until 20◦ of rotation, and for NNAR, the model fails to classify pixels to its correct label for maximum cases.