Extraction of a Specific Land-Cover Class from Very High Spatial Resolution Imagery Using Positive and Unlabeled Learning with Convolutional Neural Networks

Abstract

In remote sensing, supervised multiclass classifiers show a very promising performance in terms of classification accuracy. However, they require that all classes, in the study area, are labeled. In many applications, users may only be interested in specific land classes. When considering only one class, this referred to as One-Class classification (OC) problem. In this paper, we investigated the possibility of using Convolutional Neural Networks (CNN) within the Positive and Unlabeled Learning (PUL) framework for estimating the urban tree canopy coverage from very high spatial resolution aerial imagery. We also compared the proposed approach to the Binary CNN classification and to ensemble classifications based on various color-texture based features. The obtained classification accuracies show that PUL strategies provide competitive extraction results, especially the proposed CNN based one, due to the fact that PUL is a positive-unlabeled method in which large amounts of available unlabeled samples is incorporated into the training phase, allowing the classifier to model effectively the tree class.