Weakly supervised fully convolutional networks using OBIA classification output

Abstract

In this paper, we present a methodology for preparing reference data to be used in training a fully convolutional neural network. It is a laborious and timeconsuming task to prepare adequate training data for a Fully convolutional network (FCN) since the tiles need to be fully labeled. Weakly supervised learning is used when there are inadequate, inaccurate or incomplete training labels. In this paper, coarse labels are prepared using visual image interpretation and used to train an existing semi-automatic Object-based image analysis (OBIA) chain and subsequently used to classify a very high-resolution aerial (VHR) imagery of the city of Goma, the Democratic Republic of Congo. After accuracy assessment, fully labeled training tiles are automatically extracted and used to train an FCN designed with a skip architecture and dilated convolutions. An overall accuracy of 90.7% is attained from the tests, which demonstrates that FCN is robust to noisy labels. Future steps will entail the evaluation of ensemble voting and class probabilities in preparing the training data. This approach is promising and can address the challenge of preparing a large amount of training data for training FCNs.