INVESTIGATING STRATEGIES FOR DETERMINING THE OPTIMAL INPUT DATA AMOUNT FOR OBJECT RECOGNITION IN SATELLITE IMAGERY
DOI:
https://doi.org/10.34185/1991-7848.itmm.2023.01.087Keywords:
neural network, image recognition, satellite imagery, data processing, data augmenting, artificial intelligence.Abstract
Satellite imagery recognition is a vital application of computer vision, with numerous potential use cases in fields such as disaster management, land use, and urban planning. This study aims to identify the optimal amount of input data required for training the neural network for satellite imagery recognition. To this end, a series of rigorous experiments are conducted to investigate the impact of input data quantity on several performance metrics, including accuracy, convergence, and model generalization. The findings reveal that an increase in input data quantity generally improves model performance up to a saturation point, beyond which further increases do not lead to significant improvements and may even lead to overfitting. The research proposes several strategies for identifying this saturation point and mitigating the effects of overfitting. The results gained from this research may help in designing more effective models for satellite imagery recognition.
References
Cubuk, Ekin D. and Zoph, Barret and Mane, Dandelion and Vasudevan, Vijay and Le, Quoc V. AutoAugment: Learning Augmentation Strategies From Data. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019,
pp. 113-123.
Xie, Qizhe and Dai, Zihang and Hovy, Eduard and Luong, Thang and Le, Quoc. Unsupervised Data Augmentation for Consistency Training. Advances in Neural Information Processing Systems, 2020, pp. 6256-6268.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In Advances in neural information processing systems, pages 2234–2242, 2016.
