A Weakly-Supervised Change Detection for Multispectral Earth Observation Imagery using a Long Short- Term Memory Classifier with a Virtual Training Data Neural Generator

Abstract

This paper proposes a novel approach to improve accuracy of weakly-supervised change detection for multispectral Earth Observation (EO) imagery. The method is based on the idea to use an initial small-size EO labeled dataset to generate a larger set of virtual data. We have considered two variants of virtual data-generators based on the general architecture called Generative Adversarial Network (GAN): MLPGAN and LSGAN. The resulting virtual dataset is used to train a simple Long Short-Term Memory (LSTM) classifier. The proposed method is evaluated using the Mexico dataset acquired by the Thematic Mapper (TM) sensor of the Landsat 5 satellite. For each acquisition date, two spectral bands are considered (B4, B5). The two images have been acquired in April 2000 and May 2002, respectively. We have evaluated the change detection performances (OA, Kappa, MAR, and FAR) using two virtual data generators corresponding to considered GAN architectures. As a benchmark method, we have considered the case when the LSTM classifier is trained with the original small-size dataset without synthetic data generation.