Decision-Tree-Ensembled Quality Prediction of Vae-Based Compression for Satellite Image Transmission

This paper proposes a lightweight prediction paradigm for satellite image compression based on Variational Autoencoder, addressing the strong correlation between multidimensional remote sensing image features and hyperparameters of compression model. Ensemble learning methods based on decision trees are confirmed having better performance in terms of the trade-off between accuracy and efficiency. Additionally, feature importance-driven hyperparameter optimization framework is introduced into our works, which leverages feature importance analysis to identify key factors affecting satellite image compression and transmission. Well prediction made before the compression process can guide the selection of appropriate compression parameters. Experimental results using prediction model show excellent performance: for compression ratio (bit per pixel), the model achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.9953\left(\mathrm{R}^{2}\right)$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.0048(\text{MSE})$</tex>; for image quality (PSNR) achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.9951\left(\mathrm{R}^{2}\right)$</tex>, 0.0936(MSE). By accurately predicting these factors, this framework optimizes the bandwidth required for satellite image transmission, thereby reducing transmission costs.