Biophysical Parameter Estimation Using Earth Observation Data in a Multi-Sensor Data Fusion Approach: CycleGAN

Abstract

Water management and up-to-date soil moisture (SM) information are crucial to ensure agricultural activities in dry-land farming regions. In this context, remote sensing imagery coupled with machine learning techniques can provide large scale SM information if there is enough data for training, which is really limited in reality. In this paper, we explored the potential of cycle-consistent Generative Adversarial Network (GAN) for data augmentation for training machine learning algorithms, which try to model spatial and temporal dependencies between the SM prediction (output) and the remote sensing imagery (input features). Specifically, the freely available SAR (Sentinel-1) and optical (Sentinel-2) time series data were evaluated together to predict SM using GANs. The experiments demonstrate that the proposed methodology outperforms the compared state-of-the-art methods if there is not enough data to train a regression convolutional neural networks (CNN) to predict SM content.