| | SLO | ENG | Cookies and privacy

Bigger font | Smaller font
First page > Show document

Show document Help

Title:Metoda za prostorsko-časovno semantično segmentacijo rabe tal nad satelitskimi posnetki z uporabo grafovskih nevronskih mrež : doktorska disertacija
Authors:ID Kavran, Domen (Author)
ID Lukač, Niko (Mentor) More about this mentor... New window
Files:.pdf DOK_Kavran_Domen_2026.pdf (53,94 MB)
MD5: 01492FFDBAF6F0AAB494266570828C24
 
Language:Slovenian
Work type:Doctoral dissertation
Typology:2.08 - Doctoral Dissertation
Organization:FERI - Faculty of Electrical Engineering and Computer Science
Abstract:V doktorski disertaciji je predstavljena prostorsko-časovna semantična segmentacija rabe tal nad satelitskimi posnetki. Obstoječe metode trenutnega stanja tehnike se soočajo z omejitvami pri upoštevanju prostorskega konteksta in časovnih informacij ter s pomanjkanjem prilagodljivosti na spremembe strukture vhodnih podatkov. Predlagana metoda rešuje te izzive z uporabo grafovskih nevronskih mrež nad časovnimi vrstami večspektralnih posnetkov Zemljinega površja. V prvem koraku predlagane metode se individualni posnetki segmentirajo na regije, nato pa se oblikuje graf s prostorskimi in časovno usmerjenimi povezavami med regijami. Za vsako klasificirano regijo oziroma tarčno vozlišče se tvori usmerjen podgraf, ki vključuje sosednja vozlišča z vzpostavljenimi prostorskimi in časovno usmerjenimi povezavami, ki vodijo do tarčnega vozlišča. Sosednja vozlišča tako predstavljajo prostorsko in časovno okolico. Podgraf se posreduje v cevovod za klasifikacijo tarčnega vozlišča, kjer konvolucijska nevronska mreža najprej izračuna visokonivojske značilke nad omejitvenimi okvirji regij vseh vozlišč podgrafa. Temu sledi posredovanje podgrafa z izračunanimi značilkami v grafovsko nevronsko mrežo, ki klasificira tarčno vozlišče. Opisani postopek je izveden za vsako vozlišče oziroma regijo. Izhod predlagane metode je napovedana časovna vrsta semantično segmentiranega vhodnega območja v obliki segmentacijskih map z oznakami rabe tal. Predlagana metoda semantične segmentacije je bila preizkušena nad podatkovno zbirko DynamicEarthNet, ki vključuje časovne vrste dnevnih satelitskih posnetkov s 75 geografskih območij po vsem svetu, in primerjana s trenutnim stanjem tehnike – metodami za učenje temeljnih modelov daljinskega zaznavanja: GASSL, SeCo, SatMAE in TOV. Predlagana metoda je dosegla statistično značilno najboljše povprečne rezultate v primerjavi z metodami trenutnega stanja tehnike, in sicer povprečni mIoU 0,4145 ± 0,0051 (p-vrednosti: 0,0111, 0,0087, 2 × 10−6 in 9 × 10−5) pri uporabi prostorske okolice v podgrafih in mF1 0,5202 ± 0,0103 (p-vrednosti: 0,0287, 0,0274, 6 × 10−8 in 9 × 10−5) pri uporabi časovne okolice v podgrafih. Najuspešnejša metoda trenutnega stanja tehnike GASSL je dosegla slabše povprečne rezultate, in sicer povprečni mIoU 0,3823 ± 0,0156 in mF1 0,4908 ± 0,0198. Rezultati tako kažejo, da predlagana metoda semantične segmentacije v povprečju dosega višjo uspešnost od najuspešnejše metode trenutnega stanja tehnike za +0,0322 po metriki mIoU ter za +0,0294 po metriki mF1. Obsežna statistična analiza rezultatov je pokazala, da predlagana metoda po metrikah mF1 in mIoU dosega statistično značilno boljše rezultate od obstoječih metod trenutnega stanja tehnike. Najbolje naučen model predlagane metode je z uporabo časovne okolice v podgrafih na podatkovni zbirki DynamicEarthNet dosegel rezultat wF1 0,6905. Najvišja klasifikacijska uspešnost je bila dosežena za razred »Voda«, in sicer F1 0,9159 in IoU 0,8449. Med preostalimi petimi klasifikacijskimi razredi sta izstopala še razred »Gozd in druga vegetacija« z F1 0,7890 in IoU 0,6515 ter razred »Prst« z F1 0,6561 in IoU 0,4882.
Keywords:prostorsko-časovno, semantična segmentacija, raba tal, satelitski posnetek, grafovska nevronska mreža, temeljni model, globoko učenje
Place of publishing:Maribor
Place of performance:Maribor
Publisher:[D. Kavran]
Year of publishing:2026
Number of pages:XIV, 141 f.
PID:20.500.12556/DKUM-95633 New window
UDC:502.521:551.501.8(043.3)
COBISS.SI-ID:276911107 New window
Publication date in DKUM:23.04.2026
Views:172
Downloads:33
Metadata:XML DC-XML DC-RDF
Categories:KTFMB - FERI
:
Copy citation
  
Average score:(0 votes)
Your score:Voting is allowed only for logged in users.
Share:Bookmark and Share



Hover the mouse pointer over a document title to show the abstract or click on the title to get all document metadata.

Licences

License:CC BY-NC-ND 4.0, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International
Link:http://creativecommons.org/licenses/by-nc-nd/4.0/
Description:The most restrictive Creative Commons license. This only allows people to download and share the work for no commercial gain and for no other purposes.
Licensing start date:02.10.2025

Secondary language

Language:English
Title:Method for spatiotemporal semantic segmentation of land use on satellite imagery using graph neural networks
Abstract:In the doctoral dissertation, spatiotemporal semantic segmentation of land use on satellite imagery is presented. Current state-of-the-art methods exhibit limitations in the usage of spatial context and temporal information and additionally lack adaptability to changes in the structure of input data. The proposed method overcomes these challenges by utilizing graph neural networks on time series of multispectral images of the Earth’s surface. In the first step of the proposed method, individual images are segmented into regions, followed by the construction of a graph with spatial and temporally directed edges between the regions. For each classified region, or target node, a directed subgraph is created, which includes neighboring nodes with established spatial and temporal connections leading to the target node. These neighboring nodes thus represent the spatial and temporal neighborhood. The subgraph is then passed into the target node classification pipeline, where a convolutional neural network first extracts high-level features from the bounding boxes of the regions of all subgraph nodes. Subsequently, the subgraph, enriched with these features, is processed by a graph neural network, which performs the classification of the target node. This procedure is carried out for each node or region. The result of the proposed method is a predicted time series of the semantically segmented input area in the form of segmentation maps with land use labels. The proposed semantic segmentation method was evaluated on the DynamicEarthNet dataset, which contains time series of daily satellite imagery from 75 geographic regions worldwide, and compared against state-of-the-art methods for training remote sensing foundation models, namely GASSL, SeCo, SatMAE, and TOV. The proposed method achieved the best average results compared to the state-of-the-art methods, with statistically significant differences, reaching an average mIoU 0.4145 ± 0.0051 (p-values: 0.0111, 0.0087, 2 × 10−6, and 9 × 10−5) when using spatial neighborhood in the subgraphs, and an mF1 0.5202 ± 0.0103 (p-values: 0.0287, 0.0274, 6 × 10−8, and 9 × 10−5) when using temporal neighborhood. By contrast, the best-performing state-of-the-art method GASSL yielded inferior average results, with an average mIoU 0.3823 ± 0.0156 and mF1 0.4908 ± 0.0198. The results thus demonstrate that the proposed semantic segmentation method achieves a better average performance, compared to the best-performing state-of-the-art method, with improvements of +0.0322 in mIoU and +0.0294 in mF1. An extensive statistical analysis of the results confirmed that the proposed method achieves statistically significantly better results, compared to existing state-of-the-art methods in terms of the mF1 and mIoU metrics. The best-trained model of the proposed method achieved a wF1 score of 0.6905 on the DynamicEarthNet dataset when using temporal neighborhood in the subgraphs. The highest classification performance was obtained for the class »Water«, with an F1 score of 0.9159 and an IoU of 0.8449. Among the remaining five classification classes, the »Forest and other vegetation« class stood out with an F1 score of 0.7890 and an IoU of 0.6515, while the »Soil« class was classified with an F1 score of 0.6561 and an IoU of 0.4882.
Keywords:spatiotemporal, semantic segmentation, land use, satellite image, graph neural network, foundation model, deep learning


Comments

Leave comment

You must log in to leave a comment.

Comments (0)
0 - 0 / 0
 
There are no comments!

Back
Logos of partners University of Maribor University of Ljubljana University of Primorska University of Nova Gorica