Résumé: Pansharpening strives to improve the spatial resolution of multi-spectral images while maintaining spectral fidelity. However, existing methods usually cannot guarantee a balance between spatial and spectral quality, and degrade when handling the desirable scale of full resolution. To address these challenges, this prospective study proposes a scale-transfer learning framework via estimating spectral observation model. Specifically, we design a cross-spectral transfer network to learn an expected spectral observation model by the cycle adversarial between spectral degradation and interpolation, which describes the accurate nonlinear mapping process from multi-spectral to panchromatic images. Having the favorable spectral observation model established, a scale-transfer pansharpening paradigm with co-learning of full and reduced resolutions can be constructed. First, we develop a spectral observation model-based spatial fidelity term at the reduced-resolution scale, which can alleviate the imbalance problem of spectral and spatial information widespread in current supervised paradigms. Second, we explore the reprojection regularization from full to reduced resolution based on the spectral observation model, which facilitates the ability of the pansharpening model to be extended to the scale of full resolution. Extensive experiments demonstrate the advantage of our method over the current state-of-the-arts in terms of information balance and scale transformation. We further apply our method to produce the high-resolution normalized difference vegetation index and achieve vegetation enhancement with competitive performance. Moreover, our method is lightweight and faster than other comparative methods.

Langue: Anglais
Thème Informatique

Mots clés:
Pansharpening
Scale-transfer
Cross-spectral transfer
Spectral observation model
Reprojection regularization