Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/4D2C3DF5-631A-4924-B68C-8D0DC46CB483 | ||
| Year | 2021 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Ilias Tsichlis, "Deep learning techniques for multispectral satellite image analysis", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2021 https://doi.org/10.26233/heallink.tuc.88855 | ||
| Appears in Collections |
In the field of Machine Learning, the need to understand the operation of deep neural networks is becoming more and more important. It is well established that researchers’ attention focuses mainly on the dataset used during neural network training, as well as on the quality of the output they produce. However, this focus has not been utilized specifically for better understanding of their internal operation. The research in this diploma thesis aims to determine how this focus can improve the knowledge of how deep neural networks operate. Specifically, emphasis is placed on the architectures of two well-known and widely-used neural networks, VGG16 and ResNet50. In this context, to test the hypothesis that neural networks can be understood macroscopically, this thesis examines the creation of an appropriate, dynamic dataset for training the two networks, but also the identification of the key correlations that determine the classification decision. Specifically, a programming tool has been developed that has the ability to dynamically export satellite images from the OpenStreetMap database to form training sets for the problem of correctly classifying images containing elements of renewable energy sources (photovoltaics, wind turbines, dams). Additionally, heatmaps were used to explain the operation of the neural networks trained with these datasets in conjunction with a novel performance metric, which seems to be appropriate for the purpose of a deeper investigation into neural network classifications. A key finding from this work is that, during training and classification, certain proper or improper correlations between input and target characteristics are being developed. Our emphasis is placed on whether this information can be useful, for choosing an appropriate classifier to solve specific problems. The results of this work indicate that research, evolution, and selection of a deep neural network architecture to solve modern classification problems are choices that can be guided using information extracted according to the mentioned key finding.
Copyright © DIAS 2013
