With the rapid development of advanced technologies, especially deep learning, we urgently need a large-scale dataset supporting deeper SAR image interpretation. Currently, we present two public large-scale datasets OpenSARShip and OpenSARUrban dedicated to deeper interpretation of SAR Ship Imagery and urban Imagery respectively. Both of data are collected from Sentinel-1 Imagery. The OpenSARShip dataset consists of 34528 SAR ship chips with automatic identification system (AIS) information covering various environmental conditions. The OpenSARUrban collection provides 33358 SAR image patches of urban scenes, covering 21 major cities of China, including 10 different target area categories, 4 kinds of data formats, 2 kinds of polarization modes. These benchmarks are provided for the remote sensing community that enables extensive evaluation and investigation of deep SAR image interpretation.

We have developed novel ship detectors from SAR and Optical Satellite Images based on deep learning. For arbitrary-oriented ship detection from optical satellite images, we presented rotated region proposal and discrimination Convolutional Neural Networks to deal with this task. We also developed a coupled CNN for small and densely clustered SAR ship detection.

It is extreme labor expensive to collect such pixel-wise annotated training data for semantic segmentation and object extraction for Remote Sensing images. We intend to develop weakly supervised learning techniques to address such challenge. Recently, we introduce a label noise-adaptive (NA) fully convolution neural network trained on noisy labeled data generated by GIS building footprint data.

The overall objective of this project is to provide an effective solution for large-scale dense SITS analysis, being capable of automatic discovery of regularities, relationships,and dynamic evolution, leading to a better and easier understanding of the underlying processes ofspecific scenes and targets. The scientists of CEOSpaceTech - the research center within Politehnica University Bucharest - Romania,Tongji University and Shanghai Jiaotong University - China will tightly collaborate to advance these innovative techniques.

Deep learning have gains great success for SAR Image interpretation tasks but they are black box models the could not tell what they learn from data, how it makes decision, etc. To make more physically interpretable and transparent models, recently we propose a contrastive-regulated convolutional neural network (CNN) in the complex domain, leading to learn a physically interpretable deep learning model directly from the original backscattered data.

The miniature UAV SAR imaging system is a portable, efficient, high resolution SAR imaging system, which is composed of minisar system, UAV flight system and differential base station position system.The system has four features: 1) single-polarization and full-polarization mode, which can obtain more object information;2) light weight and easy to carry;3) work in all-weather and dark condition and improve work efficiency;4) Simple procedure and perfect software equipment.