Abstract: 【Objective】It is an effective technical method to obtain land cover information by remote sensing classification. Objective and reasonable evaluation of the uncertainty in remote sensing classification maps is significant to agricultural resources survey, crop yield assessment and other remote sensing applications. This study focuses on the problem of uncertainty evaluation for remote sensing image classification at the land cover scale. A new uncertainty evaluation criteria, boundary region-based modified rough entropy model (BMREM) is proposed based on the improvement of modified rough entropy model (MREM). 【Method】 The traditional MREM model measures the overall average uncertainty of the same class feature. But the classification uncertainty is mainly caused by the boundary between different feature classes in RS image. So, in this study, all pixels value range for class feature is replaced by the boundary range of one feature class in image. The rationality of the BMREM model is proved based on rough set theory. And then the BMREM model was established under the impact of boundary region on the uncertainty assessment in remote sensing image classification. Afterwards, the MREM model and the BMREM model are applied to evaluate the uncertainty of remote sensing classification results for one Landsat TM image and one IKONOS image, respectively. The theoretical conclusion was confirmed by uncertainty evaluation result using experimental data of different spatial resolution and different research areas.【Result】 Compared with the MREM model，the uncertainty caused by the classification knowledge can be better described using the BMREM model. So the uncertainty of remote sensing image classification is more reasonable and objective. The empirical results reveal that the irrationality of uncertainty assessment is not obvious using MREM model. It also can objectively measure the uncertainty problem of remote sensing image classification if the mixed pixel is more serious. But it is difficult to objectively evaluate the uncertainty of remote sensing image classification by MREM model if the ground objects distribution is more concentrated in study area and the area is relatively large.【Conclusion】This study proved that MREM model magnified the value of uncertainty caused by existence of boundary region from the theory analysis and model application. Yet the uncertainty generated by classification knowledge can be better depicted using the BMREM model. So the BMREM model measured the classification uncertainty at the scale of land cover classes more reasonably and accurately.

Key words: remote sensing image , classification , uncertainty , modified rough entropy , boundary region