This innovation area aims to develop deep learning models and applications for monitoring and prediction of objects, hazard risks and streamlining of aerial surveys.
Optical images from drones or satellites and data captured by radar sensors from above contain enormous amounts of complex data. They have the potential to reveal valuable information about our planet and its surface that could be used automate terrain mapping or to predict objects and hazard risks such as vessels and potential oil spills at sea.
Visual Intelligence researchers have developed novel methods for improving the monitoring and prediction of hazard risks, object detection, and for surveying and mapping the ground and sea from the air by exploiting remote sensing images. Such methods include new approaches for:
The latter two methods were developed in close collaboration with user partner Kongsberg Satellite Services (KSAT).
Limited and inadequate training data is a general problem in remote sensing. Combinations of multi-sensor data, e.g. from optical and radar sensors, and time dependencies is another key challenge. Modelling of contextual information may also enhance the performance. The methods mentioned earlier address these research challenges in different ways.
For instance, the proposed oil spill detection method is a step towards achieving uncertainty quantification in deep learning models for remote sensing data analysis.
As for all image analysis applications, the development of deep learning methodology to solve certain tasks in earth observation often benefits from solutions developed to solve other problems.For instance, when developing methods for building segmentation, ideas from segmentation of oil spills can potentially be transferred.
It can also be valuable to reveal any different behaviour of segmentation algorithms due to different properties in data sources. Aerial imagery and satellite imagery come with different resolution, contrast, noise properties, and by contrasting seemingly similar deep learning methods, the influence of different data properties may be revealed and better understood.
We have developed new methods and obtained insights into how one can use self-supervised learning when there are images acquired at two or more times available, both as a pre-training step and as an integral part of change detection. The insights and experience gained here on the concept of self-supervised learning in general has a wide range of other relevant applications. We have explored similar methodology within seismic analysis, where it has been utilized to identify and characterize geological regions in vast seismic datasets.