Change detection-aided single linear prediction of multi-temporal satellite images

Significant temporal correlation of multi-temporal remote sensed images presents an opportunity to use historical images for predicting the recent image. This paper investigates the possibilities of predicting the current image from a reference previous image using linear regression after change detection. Single linear regression model for temporal prediction is usually used for the images that are mainly affected by system or environmental noise. When the images do experience small portions of real changes, a PCA based outlier removal technique for generating a prediction model to best fit the majority of unchanged data is investigated. The PCA based change detection method first excludes the changed pixels and then the prediction method is applied on rest of the pixels. The potential of this sequential data compression is dependent mainly on multi-temporal image analysis for prediction or forecasting purposes. The transmission load can be substantially reduced by properly exploiting the temporal correlation. But if the amount of real land-cover change is not limited then the degraded temporal dependency can affect the model parameters of the prediction. The proposed change detection-aided prediction method will substantially improve the accuracy as the changed pixels have been removed.