Exploring Label Probability Sequence to Robustly Learn Deep Convolutional Neural Networks for Road Extraction With Noisy Datasets

Abstract

Recently, crowdsourced geographic data have provided a cost-effective approach to learn deep convolutional neural networks (DCNNs) for road extraction from remote sensing images. However, crowdsourced datasets often suffer from label noise and include error labels that can affect the performance of DCNN-based road extraction methods. Thus, we propose a novel sequence deep learning (SDL) framework that introduces label probability sequence to correct error labels and robustly learn DCNNs for road extraction. The label probability sequence is obtained by front-end developed DCNNs, which can provide valuable true label information to alleviate the adverse effect of label noise. Specifically, to pursue high effectiveness of the label correction process, an adaptive label correction method is proposed to exploit the distance between the true label distribution hidden in label probability sequence and the possible true label distribution. In addition, based on the uncertainty of label probability sequence, a noise correction loss function is introduced to robustly learn DCNNs on the improved dataset produced by the adaptive label correction. By repeating these two steps in each epoch, we can learn efficiently DCNN-based road extraction methods under noisy datasets. Experimental studies on two real-world noisy road datasets show that the proposed method can help to overcome label noise issue for DCNNs and achieve excellent performance in road extraction.