Razvoj modela utemeljenog na dubokim neuronskim mrežama za potrebe automatske klasifikacije satelitskih slika

master's thesis

• Title (english): Development of a deep neural network-based model for automatic classification of satellite images
• Author: Josip Prpić
• Mentor: Damir Pintar (mentor)
• Committee member: Damir Pintar (predsjednik povjerenstva)
• Committee member: Mihaela Vranić (član povjerenstva)
• Committee member: Luka Humski (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-01, Croatia

Abstract

Ovaj diplomski rad istražuje primjenu dubokog učenja u klasifikaciji satelitskih slika korištenjem modela ResNet50 i VGG19 na EuroSat skupu podataka, ekstrahiran iz Sentinel-2 skupa podataka. U radu se detaljno razmatra teorijska osnova dubokog učenja, s fokusom na konvolucijske neuronske mreže kao temeljne arhitekture za obradu slika i ostalih vizualnih podataka. Implementacija modela realizirana je u okviru TensorFlow-a uz upotrebu Keras API-ja za jednostavno definiranje i treniranje modela.
Istraživanje obuhvaća povijest i opsežnu analizu ResNet50 i VGG19 modela. ResNet50, koji koristi inovativni pristup rezidualnih blokova za olakšavanje treniranja dubokih mreža, uspoređuje se s VGG19, poznatim po svojoj dubokoj arhitekturi s konvolucijskim slojevima. Eksperiment je pokazao da VGG19 postiže veću točnost u klasifikaciji slika. Sugerirajući da je njegova složenija arhitektura efikasnija u ekstrakciji detaljnih značajki.
Analiza rezultata uključuje interpretaciju matice zabune za svaki model. Pružajući uvid u metrike poput preciznost, odziva i F1-točnosti. Kroz ove metrike identificiramo VGG19 kao preferirani model za precizne i pouzdane klasifikacije, s obzirom na mogućnost primjene sustava u stvarnom svijetu.
Zaključno, ovaj rad naglašava važnost dubokog učenja u analizi velikih skupova satelitskih podataka te potiče daljnja istraživanja usmjerena na unaprjeđenje performansi sustava kroz augmentaciju i anotaciju slika i optimizaciju hiperparametara.

Abstract (english)

This paper explores the application of deep learning in satellite image classification using ResNet50 and VGG19 models on the EuroSat dataset, extracted from Sentinel-2 dataset. The study extensively examines the theoretical foundations of deep learning, focusing on convolutional neural networks (CNNs) as fundamental architectures for processing images and visual data. Model implementation was carried out using TensorFlow with the Keras API for straightforward model definition and training.
The research includes a historical overview and comprehensive analysis of ResNet50 and VGG19 models. ResNet50, employing an innovative approach with residual blocks to facilitate training deep networks, is compared with VGG19, known for its deep architecture with convolutional layers. The experiment demonstrates that VGG19 achieves higher accuracy in image classification, suggesting that his more complex architecture is effective in extracting detailed features.
Results analysis involves interpreting confusion matrices for each model, providing insights into metrics such as precision, recall, and F1-score. Through these metrics, VGG19 is identified as the preferred model for precise and reliable classifications, considering practical real-world applications.
In conclusion, this study underscores the importance of deep learning in analysing large satellite datasets and encourages further research aimed at enhancing system performance through image augmentation, annotation, and hyperparameter optimization.

• Keywords: Duboko učenje; Konvolucijske neuronske mreže; ResNet50; VGG19; TensorFlow; Keras; Matrica zabune
• Keywords (english): Deep Learning; Convolutional Neural Networks; ResNet50; VGG19; TensorFlow; Keras; Confusion Matrix.
• Language: croatian
• URN:NBN: urn:nbn:hr:168:572786
• 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: Open access
• Created on: 2024-10-12 21:55:41