A image classification task on EuroSAT dataset with neural networks using Keras and TensorFlow.
This term paper deals with the analysis of the eurosat dataset. The aim is to create several different neural network architectures and train the chosen dataset on them. For the purpose of this thesis, a total of 4 neural network architectures have been selected:
- Multilayer perceptron (A0)
- Basic convolutional neural network (A1)
- Convolutional neural network with data enrichment (A2)
- Convolutional neural network with transfer learning (A3)
The EuroSAT dataset is designed for an image classification task, hence the main objective of this analysis is to compare different convolutional neural network architectures. The multilayer perceptron will be used only as a starting point for the research and no further modifications and experiments will be performed on it (hence it is denoted as A0).
Each architecture will include model creation, model training, training visualization and model evaluation. Various experiments will also be performed within the architectures using hyperparameter tuning. The individual results will be recorded in the corresponding tables. At the end of the analysis, all results will be summarized.
- keras
- tensorflow
- tensorflow_datasets
- matplotlib
- numpy
- David Poslušný