An Application for Scene Classification Using Transfer Learning
Abstract
Read abstract
Scene classification using satellite imagery has become an important task in computer vision. Deep learning (DL), a machine learning approach inspired by the human brain, is highly effective at identifying patterns and structures in images through neural networks. This study aims to evaluate the impact of transfer learning on scene classification performance using heterogeneous datasets. Two different scene datasets were analyzed using deep learning: the PatternNet (PN) dataset and a dataset from Türkiye (TR) collected from Google Earth. PN includes 38 categories of high-resolution satellite images, while the TR dataset consists of 17 categories. The ResNet18 model was applied to both datasets. The TR dataset was collected, and five experiments were conducted using combinations of PN and TR datasets. Transfer learning (TL) was applied using theImageNet dataset for pre-training. In the first experiment, 38 categories from PN were predicted, while the second experiment focused on 17 categories. In the third experiment, the TR dataset was used to improve learning by splitting it into training and validation sets. In experiments four and five, additional TR images were incorporated into training and validation to enhance performance. Results showed that when only PN data was used, accuracy reached 0.99 with reduced categories and 0.98 with the full dataset. When PN was used for training and TR for testing, accuracy dropped to 0.90 due to differences in color and structure. However, adding a small number of TR images improved results, achieving accuracies up to 0.9564. These findings indicate that even a small amount of target-domain data improves classification performance in transfer learning settings, particularly when dealing with heterogeneous image distributions.
Record information
- Authors
- Ayça Gözel; Elif Çiğdem Keleş
- Publisher
- Özgür Yayınları
- Publication
- Artificial Intelligence and Machine Learning Models
- DOI
- 10.58830/ozgur.pub1403.c5637
Related records
Ayça Gözel
Author profile linked by a source identifier. Not an IKSRO membership claim.
Read and explore →Elif Çiğdem Keleş
Author profile linked by a source identifier. Not an IKSRO membership claim.
Read and explore →