Kalman filter method for generating time-series synthetic Landsat images and their uncertainty from Landsat and MODIS observations

The Landsat program, since its commencement in 1972, has acquired millions of images of our planet. Those images are one of the most valuable Earth Observation resources for local, regional and global land surface monitoring and study due to their moderate spatial resolution and rich spectral information. However, their applications are impeded largely by their relatively low revisit frequency and cloud contamination on images. In order to improve their usability, a number of studies have been conducted to blend Landsat images with Moderate Resolution Imaging Spectroradiometer (MODIS) images to take merits of the two sensors. All blending models reported that they can predict synthetic Landsat images with various degrees of accuracy. However, only a couple of models reported that they can explicitly estimate uncertainty for their blended images. In this study, we propose a new surface reflectance blending model based on a Kalman Filter algorithm (Kalman Filter Reflectance Fusion Model - KFRFM) to predict time-series synthetic Landsat images from Landsat and MODIS images, and simultaneously to estimate uncertainty of the predicted synthetic images to quantify the quality of the synthetic images. Using the model, we predicted a time-series of 38 synthetic Landsat images with a temporal interval of 4 days for a vegetation growing season spanning about 6 months from Nadir Bi-directional Reflectance Distribution Function adjusted MODIS product (MCD43A4), and their corresponding uncertainties. From this time-series, we calculated five vegetation indices involving all the spectral bands of the synthetic images, and compared them to those from Landsat observations. The results demonstrated that the proposed method is able to produce high quality synthetic Landsat images to meet various application demands for higher spatial and temporal resolution images. Uncertainty analysis reveals that cropland has the largest uncertainty followed by grassland while forests have the smallest uncertainties among the seven vegetation land cover types of the study area. For performance evaluation, we compared KFRFM to several published models. The comparison results reveal that KFRFM performs the best based on the assessed image quality indices.