CC BY 4.0Küster, JannickJannickKüsterGroß, WolfgangWolfgangGroßSchreiner, SimonSimonSchreinerMichel, AndreasAndreasMichelHeizmann, MichaelMichaelHeizmann2025-04-232025-04-232025https://doi.org/10.24406/publica-4557https://publica.fraunhofer.de/handle/publica/48688310.24406/publica-4557This paper analyzes the characteristics of water spectra and examines how varying training datasets, with different amounts of water spectra, impact the reconstruction accuracy of lossy hyperspectral data compression methods. The hypothesis is that the spectral signatures of water have lower intensity values than the spectral signatures of land and, therefore, have fewer significant features or that the features have a lower intensity. At lower intensity values, the reconstruction error caused by the lossy compression has a greater influence on the signal, as the relative error is more significant. Furthermore, the scarcity and less pronounced features in water spectra hinder the training of machine learning methods. To validate this hypothesis, lossy compression models are trained using three distinct training datasets: a combination of land and water spectra, exclusively land spectra, and exclusively water spectra. The experimental results substantiate the hypothesis that the lower intensity and weaker features of water spectra pose significant challenges for machine learning models. However, the reconstruction accuracy of the A1D-CAE demonstrates that combining land and water spectra enables a generalized restoration of critical features, thereby mitigating the limitations of models trained exclusively on land- or water-based datasets. Consequently, it is imperative to include a sufficient amount of water spectra when constructing training datasets.enHyperspectral image (HSI) compressionspectral compressionfeature extractionautoencoderspectral analysisconference paper