| dc.contributor.advisor | Villalón-Turrubiates, Iván E. | |
| dc.contributor.advisor | Martínez-Sánchez, Víctor H. | |
| dc.contributor.author | Cordero-Robles, Carlos A. | |
| dc.date.accessioned | 2020-09-25T22:26:32Z | |
| dc.date.accessioned | 2023-03-10T18:12:25Z | |
| dc.date.available | 2020-09-25T22:26:32Z | |
| dc.date.available | 2023-03-10T18:12:25Z | |
| dc.date.issued | 2020-08 | |
| dc.identifier.citation | Cordero-Robles, C.A. (2020). Performance comparison of deep learning models applied for satellite image classification. Trabajo de obtención de grado, Maestría en Sistemas Computacionales. Tlaquepaque, Jalisco: ITESO. | es_MX |
| dc.identifier.uri | https://hdl.handle.net/20.500.12032/71790 | |
| dc.description | Satellite images classification is important for applications that involve the distribution of the human activities. Such distribution helps the governments to determine the best places to expand cities avoiding problems related to natural disasters or legal constrains. Currently, existing few agencies in charge of image classification and the area to cover is enormous. Therefor an automation of this process is necessary for this task otherwise, it will take an eternity to perform this task manually. On the other hand, detection and classification algorithms used before Machine Learning (ML) have not shown good result classifying this specific sort of images. However, latest approaches for image classification using Convolutional Neural Networks (CNN) have shown quite accurate results. In this research, we analyses the performance in four different CNN architectures used for satellite image classification. We use a dataset provided in 2017 by IARPA names IARPA fMoW. It contains more than two thousand images belonging to 62 classes already separated in train and validation.
The solution was implemented in Python using the Keras and Tensorflow libraries. The research was divided in two parts: Hyperparameters optimization and architectures results evaluation. For the first part we used only seven classes from a sample of the dataset (The sample is three hundred times smaller than the complete dataset). The architectures are trained using these seven classes of this small dataset to determine the best hyperparameters. After having selected the hyperparameters the architectures are trained with the complete sample. The evaluation is based on visual examination with the help of the tool Tensorboard and SKLearn metrics.
All the architectures showed accuracies near to 90% over the training dataset sample. The architecture with the best accuracy result was Resnet-152 with one accuracy of 99% over the training dataset Sample. The accuracy over the validation dataset will become important after training the architectures with the complete dataset. The training with the complete dataset will be performed in future works. | es_MX |
| dc.description.sponsorship | ITESO, A. C. | es |
| dc.language.iso | eng | es_MX |
| dc.publisher | ITESO | es_MX |
| dc.rights.uri | http://quijote.biblio.iteso.mx/licencias/CC-BY-NC-2.5-MX.pdf | es_MX |
| dc.subject | Deep learning | es_MX |
| dc.subject | Satellite images | es_MX |
| dc.subject | Image classification | es_MX |
| dc.title | Performance comparison of deep learning models applied for satellite image classification | es_MX |
| dc.type | info:eu-repo/semantics/masterThesis | es_MX |
