Politecnico di Torino (logo)

Feature extraction using satellite images

Omid Toutian Esfahani

Feature extraction using satellite images.

Rel. Carlo Novara. Politecnico di Torino, Corso di laurea magistrale in Mechatronic Engineering (Ingegneria Meccatronica), 2019

PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (6MB) | Preview

Remote Sensing Technology (RST) mainly focuses on the gathering information about the Earth's surface and atmosphere using sensors onboard airborne or space-borne platforms. RST has been widely used in ground mapping, resource regulation, environmental protection, urban planning, geological research, disaster relief and emergency, military reconnaissance, and many other fields. RST has enabled us to have access to a large amount of data at a relatively low cost and it would be feasible to perform computational algorithms on gathered image data to extract useful information. Amongst the applications of RST, object detection plays a key role in detecting and identifying target objects. Target object detection and identification are usually achieved using a combination of signal and image processing techniques and statistical models. Availability, accessibility to data, the evolution of Artificial Intelligence and largely increased computing power has enabled the scientists to solve many of today's world problems which were considered unsolvable before. This improvement is not only limited to a specific field of science but also proved to be useful in many different fields, varying from economics and finance to robotic and image processing. Recent advances in deep learning architectures have shown promising results over statistical counterparts in target object detection and identification. Although such learning architectures are heavily dependent on computing resources, they are easy to use compared to sophisticated statistical models. Further, deep learning architectures are also able to render feature engineering as a part of their learning process which makes them extremely powerful in target object detection and identification process. Nowadays object detection methods are maturing very rapidly and thanks to deep learning, new algorithms and models keep on outperforming the previous ones. For instance, methods such as SSD, YOLO, R-CNN, Fast R-CNN, Faster R-CNN are among the state of art models and they are able to deliver high accuracy and reasonable processing speed for a wide variety of applications. Although several breakthroughs have been witnessed in recent years in object detection, still there is room to improve. Object detection algorithms have the potential to be applied to satellite images to detect the various object. One important application of object detection on satellite images is ship detection due to increasing of shipping traffic in recent years. There are more ships used in today's world and this phenomenon leads to increase the chances of breaches at sea like ship accidents, piracy, illegal fishing, drug trafficking, and illegal cargo movement. This has urged many organizations such as environmental protection agencies, insurance companies, and national government authorities, to have a closer look and pay more attention over the open seas. The main focus of this thesis is on ship detection based on satellite images by employing Deep Learning algorithms. For the specific case of ship detection, one of the biggest challenges is the lack of appropriate dataset. Therefore, a synthetic dataset is created to overcome the issue. For object detection, various methods are implemented and compared to achieve the best result.

Relators: Carlo Novara
Academic year: 2018/19
Publication type: Electronic
Number of Pages: 73
Corso di laurea: Corso di laurea magistrale in Mechatronic Engineering (Ingegneria Meccatronica)
Classe di laurea: New organization > Master science > LM-25 - AUTOMATION ENGINEERING
Aziende collaboratrici: AIKO S.R.L.
URI: http://webthesis.biblio.polito.it/id/eprint/10938
Modify record (reserved for operators) Modify record (reserved for operators)