A new method for segmentation of a remotely sensed image is proposed. This method consists of division and merging. The image is divided into spatially uniform areas by the division. The division rule is based on the difference of the coefficient vectors of a local regression model fitted to neighboring areas of the image. If the difference is not significant, these areas are regarded as spatially uniform. As the difference is a F-statistic, a statistical test is employed to evaluate the significance of the difference, and the threshold for the test is theoretically derived. An area is divided into two subareas when the null hypothesis “coefficient vectors of a local regression model fitted to these two subareas are identical” is rejected. The division is repeated until all subareas are spatially uniform.
After the division, divided subareas are merged into uniform areas by reapplying the statistical test. The merging is based on the similarity of the coefficient vectors for the subareas. Merged areas are labeled for output.
In this paper, we describe the regression model, the formulation for the statistical test, and the determination of the threshold. The validity of this method is confirmed by numerical simulation. The processing result for an actual remotely sensed image is also shown.