xAI Short Description

Explainable Deep Learning for Earth Observation

Artificial Intelligence (AI) is the epicenter of a new revolution boosted by Machine and Deep Learning (DL) methods, computational power, and fostered by Big Data. Big Data is the process of collecting large volumes of data via IT systems, sensors, in-situ observations, people, or IoT, only to enumerate a few modalities. AI is the convergence of theoretical methods and tools to transform Big Data in useful information and knowledge for making decisions, verifying hypotheses, understanding insights, or making predictions.

Satellites are the only global Earth Observation (EO) data source. The field of EO is presently at a key turning point, Big EO Data are now freely and openly accessible. Within the frame of the Copernicus program [1] alone, almost 30 million EO Sentinel satellite products have been generated recently and downloaded by ca. 350,000 users.

AI is currently studied mainly for optical imagery, i.e. photography. EO images are basically different and much more complex. AI for EO requires specific methods for the full information extraction from spatial, temporal or spectral information at a global scale. This involves new paradigms to analyze jointly multimodal sensor records as the EO multi-sensor data optical, IR or microwaves.  EO records data of high complexity, physically-based, dynamic, non-linear coupled Earth System. We need to develop new AI paradigms with integrated physical principles into the learning mechanism. These are well beyond and do not emerge from the present cats and dogs recognition techniques. Thus, there is a huge motivation in developing AI for EO methods and exploiting the results.

This context and new tendencies show us the directions where we should go. Big EO Data valorization is our use case for the elaboration of new specialized AI methodology for EO sensor data, Explainable Deep Learning for Earth Observation: xAI Deep Sensing.

We address new and very specific topics in EO for learning scene semantics, scene signatures, and physical parameters. Therefore, we develop new DNN architectures and hybrid paradigms - encapsulating physical knowledge and designing appropriate training data sets and procedures supported by consensus and consistency in applications. 

OBJECTIVES

xAI Deep Sensing overall objective is a major development in AI for EO, a ground-breaking paradigm shift for maximization of information extracted from EO data, aiming to strengthen and broaden the EO value chain for applications and markets, that is to realize a new paradigm and a breakthrough for cohesive and uniform EO information exploration and valorization based on the existing data EO missions demonstrating Data Mining and Virtual Sensing methodologies. This is an EO intelligence paradigm change from data to information and knowledge valorization. In the context of the presented necessities and tendencies, xAI Deep Sensing has 5 principal research objectives:

O1. Elaborate self-learning AI: develop unsupervised learning for EO data without or with very few labels, and transfer learning for multi-sensor EO and cross modalities generalization (Task. 2.1)

O2. Elaborate physics-aware AI: integrate physical principles and domain knowledge with learning and statistical methods, exploit synergy of models and methods (Task 2.2)

O3. Elaborate explainable AI methods for EO: hybrid paradigms to make transparent the AI black box, quantify errors and uncertainty, make transparent the estimation and decision processes, provide confidence measures for the results (Task 2.3)

O4. Demonstrate the results to AI and EO stakeholders and practitioners as functionalities of Data Mining for decision making (Task 3.1) and Virtual Sensing prediction processes (Task 3.2).

O5. Disseminate the project results via articles in highly recognized international journals and conferences and include particular topics in UPB MS curricula. We envisage publishing 6 journal papers (3 under Gold standard open access) and 6 papers in conferences.

ACKNOWLEDGEMENT
This work was supported by the grant of the Romanian Ministry of Education and Research, CNCS - UEFISCDI, project number PN-III-P4-ID-PCE-2020-2120, within PNCDI III.