Atıf İçin Kopyala
Gharagoz N. S., Karaca A. C.
Applied and Computational Engineering, cilt.100, sa.1, ss.153-161, 2025 (Hakemli Dergi)
Özet
Multispectral (MS) satellites generate a large volume of data and compression of this data is a critical task for remote sensing applications. Multitemporal images, captured at different dates over the same scene, contain temporal correlations that can be used for compression. This study is built on MultiTempGAN, a lightweight generative adversarial network (GAN) designed for multitemporal MS image compression, and investigates the impact of post-training model quantization on compression performance. By reducing the precision of model parameters to various bit-widths, quantization significantly reduces the model size while preserving image quality. Experiments conducted on Sentinel-2 satellites MS image pairs show that the quantized models achieve similar signal-to-noise ratio (SNR) and bit-per-pixel (bpp) metrics as the original model, with minimal impact on image reconstruction and compression efficiency. In addition, the reduction in model size facilitates more resource-efficient deployment, supporting large-scale remote sensing applications. These findings highlight the potential of model quantization to optimize deep learning-based compression techniques, enabling scalable and efficient handling of MS data without sacrificing accuracy and performance.