Deep-Learning Applications for Time Series SAR (Sentinel-1) Image Classification with Limited Labelled Area

Deep-learning is an emerging technique in various fields and demonstrating an outstanding performances combined with big data. It is also true for remote sensing field where the number of available satellite images are increasing along with improved spatio-temporal resolution. However, since deep-learning requires vast amount of learning materials for desirable performance, the lack of labelled data is a limiting factor in some fields. Especially, the imbalance between labelled and unlabeled data is dramatic in remote sensing since the satellite images are acquired in near-real time while accurate labeling for those images takes times. In this study, diverse applications such as data augmentation, semi-supervised classification, domain adapted architecture were performed to overcome the limitation of lacking labeling data. The applications were examined on rice paddy classification in South Korea as it can fully utilize the advantage of non-linear modelling of deep-learning and the images of high spatio-temporal resolution. The deep-learning model is trained with confined area of Dang-Jin city and applied to the entire nation. After then, all possible combination of the three application were evaluated with pixel based comparison in ten sites and city-level comparison using national statistics. The accuracy and kappa value calculated from pixel based evaluation have increased in all applications. Especially, the data augmentation greatly increased in accuracy of 1.5% from 94.91% to 96.42%. The improvement was followed by semi-supervised classification and domain adapted architecture that increased 1.02% and 0.54% each. In the combination of the applications, it tends to follow the better performance between the combined applications. Evaluation with national statistics also recorded the best R² at the data augmentation with 0.96 and showed little difference in domain adapted architecture. The results imply that the aforementioned applications can contribute to increase model performance, even though they need careful attention in the process of using or imitating unlabeled data. The results of this study can enhance the applicability of deep-learning in remote sensing field, especially when the target class has high intra-class heterogeneity that needs sufficient labelled data to train it.