Data assimilation is a strategy to foster data integration and data harmonisation in a bi-directional way between the measured and the modelled reality. In other words, it aims to combine measurements (observations) with the understanding of the spatio-temporal properties and evolution of system’s variables or properties and model information about them. Models can be calibrated and keeping them ‘on track’ by constraining them with observations. Vice versa, observations can be validated through models. Approached as a mathematical problem, data assimilation aims at minimizing cost functions or penalize a function to ensure optimality in fitting. Equations are used to describe system parameters and the relationships among them, It is noteworthy, that models encompass information from previous measurements, experiences, and theory. While the observations are influenced by (known) properties such as precisions, etc. of the measurement devices, the robustness of models rely on the consolidated knowledge. Because uncertainties reside in all components with unknown or even undeterminable errors, the approach is usually probabilistic, including Bayesian and other related techniques. Widely used in meteorological sciences, successful data assimilation has been boosted the reliability of weather forecast , while sensitivity to errors remains.
In Earth observation, data assimilation compensates for the fact that a specific site could be observed in a variety of measurements by satellites with different sensor types, at different dates, different angular geometries and viewing directions, illumination conditions (solar time), observation frequencies, etc. In particular, for monitoring processes, measurements over time need to assure to actually measure the status of the system or object and not the divergence in observation. To overcome these divergences and converge them with the actual properties of an observed object or target class such as spectral or geospatial properties, observation modelling can be considered an important contribution from geospatial theory. this also links to class modelling or geon modelling. The synergy of a vegetation growth model and a remote sensing observation model can be exploited to improve the retrieval of geo-biophysical information. For vegetation and crop type monitoring radiative transfer modelling (RTF) is being used as an example.
Data assimilation can also serve in bridging the gaps between non-availabilities of EO data and other observations, to provide estimates or prediction for geographical variables, testing of hypotheses or continuous observation (monitoring). A related aspect is data imputation, i.e. filling gaps in observations e.g. by other, complementary data sets (e.g. Radar imagery in the absence of VHR data in cloudy weather conditions). Recently, these sources can also be complemented by crowd mapping and citizen science.
When interpretation of data comes into play, such as image classification, we introduce another level of uncertainty. Thus the community seeks for rigorus classifiers based on solid spectral models, acting across sensors. Semantic enrichment of satellite data is a related strategy for reaching to interpreted data in a rigorous way.
Summarizing, data assimilation comprises steps to improve the level of interpretability of the input data, by enrichment (get rid of spatial/temporal gaps), by accounting for heterogeneity (through harmonization), and by integration (combination with other data that is relevant to the application). Thereby, datasets become more comparable to each other.