GeoPython 2023

Automated and robust geometric and spectral fusion of multi-sensor, multi-spectral satellite images
2023-03-06, 14:30–15:00, 01.S.21 (1st floor)

The talk presents AROSICS and SpecHomo, two open-source and easy-to-use Python packages for automated and robust geometric and spectral fusion of multi-sensor, multi-spectral satellite images.


Earth observation satellite data acquired in recent years and decades provide an ideal data basis for accurate long-term monitoring and mapping of the Earth's surface and atmosphere. However, the vast diversity of different sensor characteristics often prevents synergetic use. Hence, there is an urgent need to combine heterogeneous multi-sensor data to generate geometrically and spectrally harmonized time series of analysis-ready satellite data. In this context, we present two open-source Python algorithms for geometric and spectral harmonization of multi-sensor, multi-temporal satellite data. AROSICS, a novel algorithm for multi-sensor image co-registration and geometric harmonization, detects and corrects positional shifts between two input images and aligns the data to a common coordinate grid. At its core, the algorithm is based on phase correlation in the frequency domain, which makes it particularly robust to spectral differences. The Python package automatically calculates thousands of control points per image pair, that are efficiently filtered in a three-stage outlier detection process. The second algorithm, SpecHomo, was developed to unify differing spectral sensor characteristics. One particular innovation is the use of material-specific regressors, i.e., separate transformation functions, optimized for different land cover classes. These regressors not only allow higher accuracies in the transformation of the spectral information, but also enable the estimation of unilaterally missing spectral bands. For example, spectral bands in the red edge region can be estimated from Landsat-8 using the presented Python package. Such synthesized red edge information has proven valuable when retrieving vegetation-related parameters such as burn severity. The talk illustrates the effectiveness of the developed algorithms to reduce multi-sensor, multi-temporal data inconsistencies and demonstrates the added value of geometric and spectral harmonization for subsequent products. It shows that the combination of multi-sensor time series offers great potential for more accurate monitoring and mapping of quickly evolving environmental processes.