Ecosystem assessment is an important basis for ecological problems identification, ecosystem restoration, biodiversity protection, and ecological compensation. It is also a piece of decision-making information for protecting national ecological security and promoting the construction of ecological civilization. Ecosystem assessment involves multiple aspects of the ecosystem, such as carbon sequestration, soil and water conservation, biodiversity, etc. Observation data covering multiple ecosystem services and different scales is the fundamental requirement to carry out a comprehensive assessment. Field observation and remote sensing data are the major data sources for ecosystem assessment. However, there are still many problems while directly using these data in assessment, such as different observation protocols, incomplete observation elements, insufficient time continuity, and inconsistent observational scales, which would bring large uncertainty to ecosystem assessment. How to fuse these multi-sources data and accurately extract ecological parameters is the key to effective ecosystem assessment. According to the observational scale, this paper classified the observation data into four types, including ground observation data, near-surface remote sensing data, airborne remote sensing data, and spaceborne remote sensing data, then reviewed the characteristics and limitations of these multi-source data, and summarized data fusion methods between them. Ground observation data is always discrete distribution and limited coverage and is recognized as ground truth for the remote sensing data. With ecological theory, such as allometry theory or species-area relationship, ground observation data from different sources can be standardized or fused to improve the temporal or spatial coverage of ecological parameters. Near-surface and airborne remote sensing data have fine spatial resolution and large spatial coverage, which can be directly retrieved or linked with ground observation data to produce scale-matched ecological parameters which can be considered as ground truth-like data. Spaceborne remote sensing data is spatially and temporally continuous observation and is used as an important data source for ecosystem assessment. The machine learning approaches are the conventional method in the ground observation data and spaceborne remote sensing data fusion, but it is usually suffered by the spatial mismatch between ground survey and pixel size. Using near-surface and airborne remote sensing data as a medium is a new solution in ground observation data and spaceborne remote sensing data fusion which can avoid scale mismatch problems and reduce the uncertainty of ecological parameters. Moreover, this paper used several key ecological parameters as examples, such as productivity, carbon sequestration capacity, and biodiversity, to introduce the multi-source data fusion methods proposed by the National Key R&D Program of China "Ecosystem Assessment Technology and Application Research Based on Multi-source Data Fusion". Finally, we summarized multi-source data fusion methods and pointed out directions for future research.