Can Crowdsourcing Support Remote Sensing Image Classification?

Abstract

There is great value to extract artificial surface from remote sensing images to understand urban expansion dynamics. With crowdsourcing data like Open Street Map (OSM), a great amount of labeled training data can be used as input of many supervised classification methods like Neural Network. This study explores the potential application of combining crowdsourcing data and remote sensing images in artificial surface extraction. A 1000 km2 area of a Landsat 8 image in Beijing, the capital city of China, is chosen as the case study. Comparing with a spectral method Normalized Differential Building Index (NDBI) and an unsupervised method ISODATA, the freely available labeled building foot scripts by OSM are used as training datasets for several supervised classification methods including Maximum Likelihood Classification (MLC), Supporting Vector Machine (SVM), and Neural Network (NN). The estimation by OSM point features with building-like attributes shows that the accuracies of the five classification methods NDBI, ISODATA, MLC, SVM, and NN are 8.51 %, 45.39%, 75.18%, 85.11 %, and 93.62% respectively. This means that the combination of crowdsourcing and remote sensing has a very potential value for satellites applications like artificial surface extraction.