Ekstrakcija geoprostornih podataka iz zračnih snimaka primjenom tehnika dubokog učenja

master's thesis

• Title (english): Extraction of Geospatial Data froam Aerial Images Using Deep Learning Techniques
• Author: Stjepan Bašić
• Mentor: Tomislav Hrkać (mentor)
• Committee member: Tomislav Hrkać (predsjednik povjerenstva)
• Committee member: Siniša Šegvić (član povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-11, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U ovom radu predstavljeni su rezultati te proces izrade modela dubokog učenja za semantičku segmentaciju pokrova tla u 6 klasa kojemu su ulazni podaci dronske snimke ljudskih naselja. Obrada i priprema ulaznih podataka vršena je u alatu ArcGis Pro, a krajnji model je pogodan za korištenje i primjenu u istom alatu. Ulaz u model je slika rezolucije 256x256 gdje jedan slikovni element odgovara stvarnoj površini 30x30 centimetara. Korištene klase pokrova su: Zgrada, Visoka Vegetacija, Niska Vegetacija, Otvoreno, Popločeno i Voda. Korištena arhitektura temeljena je na mreži U-Net s ResNet50 enkoderskim dijelom. Istrenirani model pokazao je vrlo dobre rezultate uz preciznost na validacijskom skupu od 95.38% te težinski prosjek IoU metrike od 0.9112.

Abstract (english)

This paper presents the results and the process of developing a deep learning model for semantic segmentation of land cover in 6 classes where human settlement drone shots are the input data. Data preprocessing and preparation was performed in ArcGis Pro software, and the final model can be deployed and used within it. The model input is a 256x256 pixel resolution image where one picture element corresponds to a real-world area of 30x30 centimeters. Land cover classes used were Building, High Vegetation, Low Vegetation, Open Terrain, Paved Area and Water Surface. The model is based on U-Net network architecture with a ResNet50 encoder. The final trained model showed promising results with validation set accuracy of 95.38% and a weighted average IoU score of 0.9112.

• Keywords: duboko učenje; U-Net; ResNet; semantička segmentacija; klasifikacija pokrova tla; zračne snimke
• Keywords (english): deep learning; U-Net; ResNet; semantic segmentation; land cover classification; areal imagery
• Language: croatian
• URN:NBN: urn:nbn:hr:168:852889
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 14:30:13