upcoming event

Symposium: Satellite Imagery Classification of man-made structures using Deep Learning/Machine Learning

Date28 October 2022
Time13:00-17:00
LocationVrije Universiteit Amsterdam IVM, De Boelelaan 1111, 1081 HV Amsterdam, Meeting room NU-8A46
Event organiserIVM, Vrije Universiteit Amsterdam

Topic 1: From algorithms to application dealing with different types of RS data/sensors

Scaling up from small to large scale applications is a challenge that has a prominent place in mapping manmade structures using remote sensing and deep learning/machine learning. The availability of high-resolution remote sensing imagery is increasing dramatically due to the recent trend of open-access earth observation archives. Yet, single satellite images cover a relatively small area which makes combining data from different times or even different sensors inevitable in order to map whole countries or even continents. This will introduce discrepancies between the images used during training and the images used while running inference. How to choose the most suitable remote sensing data and which (pre-)processing steps should be taken to increase the mapping accuracy is the first topic of the symposium.

Key words: remote sensing data, deep learning application, transfer learning

Topic 2: Beyond accuracy – Black box and interpretability challenges in deep learning-based remote sensing analysis

Deep Learning application on satellite imagery shows a steady increase due to its ability to efficiently deal with large volumes of data. However, its wide-spread embracement is not without risks of misuse due to the inherent blackbox characteristics of such method. Furthermore, the successes it achieved in scientific disciplines and commercial fields can sometimes lead to an unconditional trust. Thus, the second topic would aim to discuss the common pitfalls that deep learning poses and the ways to alleviate such issues, ultimately highlighting the key needs and challenges to steer future development towards better practices such as transparency, interpretability and explainability.

Key words: explainable AI