GeoDeep: Pushing the Limits of Deep Learning for Large Scale Remote Sensing Scene Analysis

This project started in 2018.
The project is funded by CENIIT.

People

The principal investigator for this project is Fahad Khan.

Introduction

Recent years have seen an ever-increasing amount of high spatial resolution (HSR) remote sensing data with abundant spatial and structural patterns. This large amount of remote sensing data at varying scale, temporal, and spatial resolutions is captured through a multitude of different spaceborne or airborne sensors, thereby containing rich information about our environment. Automatic extraction of semantic visual information is challenging particularly in urban areas due to factors such as, presence of many object classes within a single HSR image, heterogeneous structures with fine-grained texture details, and the strongly varying between-class dependencies. One of the main problems is that of scene classification where the task is to automatically associate a semantic class label to each HSR remote sensing image containing multiple land cover types and ground objects. The semantic classification task can be further decomposed into automatically associating a semantic class label to each object (detection) or pixel (segmentation) in images. The goal of this project is to go beyond state-of-the-art by developing novel methods for semantic classification of large scale HSR remote sensing data.

Industrial Motives

The results from the proposed project will be relevant in different uses cases, e.g. decision support or mission planning based on advanced terrain analysis is one of keys to success. Also, radio propagation requires high resolution terrain classification. Intelligent earth observation through remote sensing images, acquired from airborne sensors, is of high demand with numerous application potential, such as natural hazards detection, advanced terrain analysis, land use land cover analysis, law enforcement, environment monitoring, and urban planning.

Research Context and Project Vision

The project will be carried out at the Computer Vision Laboratory (CVL). CVL has a strong background in computer vision and participates in WASP and excellence centers: ELLIIT and CADICS. CVL has been successfully developing image and video classification, detection and segmentation methods with special emphasis on deep learning in recent years. Our objective is to utilize the prior knowledge from these domains to develop new deep learning methods for semantic classification of large scale HSR remote sensing satellite data. Our long-term vision is to be a leading international research environment in remote sensing and photogrammetry with focus on visual understanding of large scale remote sensing data from any spaceborne or airborne sensors. In addition to satellite imagery, remote sensing technology based on small unmanned airborne vehicles (sUAV, drone, RAS) is developing explosively.

Project Perspective

The scope of the project has been developed in cooperation with the industrial partner, Vricon in Linköping. Vricon is one of the leading companies serving the geospatial market with technology to accurately represent all visible objects on the earth.

Publications

1. Rao Muhammad Anwer, Fahad Shahbaz Khan, Joost van de Weijer, Matthieu Molinier, Jorma Laaksonen, "Binary Patterns Encoded Convolutional Neural Networks for Texture Recognition and Remote Sensing Scene Classification", ISPRS Journal of Photogrammetry and Remote Sensing , 138, 74-85, 2018.

2. Felix Järemo Lawin, Martin Danelljan, Fahad Shahbaz Khan, Per-Erik Forssén, Michael Felsberg, "Density Adaptive Point Set Registration", Proceedings of CVPR 2018 , IEEE Computer Society, 2018.

3. Goutam Bhat, Joakim Johnander, Martin Danelljan, Fahad Shahbaz Khan, Michael Felsberg, "Unveiling the Power of Deep Tracking", Proceedings of ECCV 2018 , Springer, 2018.

4. Abdelrahman Eldesokey, Michael Felsberg, Fahad Shahbaz Khan, "Propagating Confidences through CNNs for Sparse Data Regression", Proceedings of BMVC 2018 , BMVA, 2018.

5. Goutam Bhat, Martin Danelljan, Fahad Shahbaz Khan, Michael Felsberg, "Combining Local and Global Models for Robust Re-detection in Generic Object Tracking", Proceedings of AVSS 2018 , IEEE Computer Society, 2018.