magistrsko delo št.: 155/II. GIG
Primož Skledar (Author), Krištof Oštir (Mentor), Dejan Grigillo (Co-mentor), Jernej Nejc Dougan (Co-mentor)

Abstract

Konvolucijske nevronske mreže so v zadnjem desetletju v zelo velikem razvoju in se uporabljajo na skoraj vseh področjih znanosti. Pri nalogah prepoznave poplavnih območij se v veliki meri uporabljajo nevronske mreže, ki zagotavljajo avtomatizirano prepoznavo le teh z zanesljivimi rezultati. Ti rezultati so pomembni za ocenjevanje škode in pri načrtovanju obnove poplavnega območja. V raziskavi sem uporabil dva modela konvolucijske nevronske mreže, in sicer MobileNetV2 in dve stopnji EfficientNet. Za učenje modela sem uporabil podatke satelita Sentinel-2. Za ločevanje poplavljenega in nepopravljenega območja sem izdelal lastno zbirko oznak. Izdelali smo program, ki se uporablja za predobdelavo podatkov in učenje modelov. Uporabljene modele sem testiral s spreminjanjem hiperparametrov. Prav tako sem izvedel test spreminjanja po nivoju produkta in izbiro kanalov satelita Sentinel-2. V tretjem delu testiranj sem izboljšal rezultat zgolj z bogatenjem količine podatkov. Po vsakem testiranju sem podatke analiziral in pridobil optimiziran model, kot rezultat, ki je sposoben uspešno prepoznati poplavno območje. V izbranem GMS-GIS-u sem uporabil izdelano metodo in jo preizkusil na novih podatkih.

Keywords

geodezija;magistrska dela;GIG;strojno učenje;semantična segmentacija;konvolucijska nevronska mreža;hiperparametri;binarna klasifikacija;MobileNetV2;EfficientNet;Sentinel-2;poplave;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FGG - Faculty of Civil and Geodetic Engineering
Publisher: [P. Skledar]
UDC: 528.7:556.531/.532(043.3)
COBISS: 125106435 Link will open in a new window
Views: 77
Downloads: 23
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Detection of flooded areas from satellite imagery Sentinel-2 with convolutional neural network models
Secondary abstract: Convolutional neural networks have developed a lot over the past decade and are used in almost all fields of science. Neural networks are already used to a large extent for detecting flooded area and provide automated detection with reliable results. These are important for assessing damage and planning the reconstruction of the flood areas. In the study, I used two models of the convolutional neural network MobileNetV2 and two stages of EfficientNet. For learning models, I used Sentinel-2 satellite data. To separate the flooded and not flooded areas, I created my own collection of annotations. We have developed a program that we use to pre-process data and learn models. The models used were tested by changing the hyperparameters. I also performed a product-level change test and a selection of Sentinel-2 satellite channels. In the third part of the tests, I improved the results only by enriching the amount of data. After each test, I analyzed the data and obtained an optimized model as a result that can successfully detect the flooded area. In the selected GMS-GIS, I used the developed method and tested it on new data.
Secondary keywords: geodesy;master thesis;machine learning;semantic segmentation;convolutional neural network;hyperparameters;binary classification;MobileNetV2;EfficientNet;Sentinel-2;floods;
Type (COBISS): Master's thesis/paper
Thesis comment: Univ. v Ljubljani, Fak. za gradbeništvo in geodezijo
Pages: XI, 74 str.
ID: 16901254