Kalman filter for statistical monitoring of forest cover across sub-continental regions

Abstract

The Kalman filter is a multivariate generalization of the composite estimator which recursively combines a current direct estimate with a past estimate that is updated for expected change over time with a prediction model. The Kalman filter can estimate proportions of different cover types for sub-continental regions each year. A random sample of high-resolution satellite scenes at time t=0 is classified into detailed categories to estimate proportion of cover types in the population, and its sample covariance matrix. Coarse-resolution weather satellite data at time t=0 are also classified for the entire study area (i.e. a census), but less detailed categories are used that are consistent with the coarse resolution of these data. The Kalman filter combines the two multivariate population estimates into a composite vector estimate. A probability transition matrix can estimate the proportions of the detailed cover types at time t=2 given the previous composite estimate at time t=1. This estimate can be combined with any new estimates for time t=2 (e.g. new coarse-resolution satellite data). This procedure is repeated for each time period.

Citation

Czaplewski, Raymond L. 1991. Kalman filter for statistical monitoring of forest cover across sub-continental regions. Biometric Bulletin. 8(4): 6016.