提出了一种借助多偏移遥感图像来改进基于BP神经网络(BPNN)的亚像元定位新方法.不同于原BPNN方法使用单幅低空间分辨率观测图像, 新方法利用多幅带有亚像元偏移的低空间分辨图像来确定亚像元属于各类的概率, 然后根据概率值和地物覆盖比例确定亚像元类别, 以降低BPNN定位模型中的不确定性和误差.实验表明, 提出方法在视觉和定量评价上, 均能获得更高精度的亚像元定位结果, 验证了提出方法的有效性.

A new sub-pixel mapping method is presented in this paper, which makes use of multiple shifted remote sensing images to enhance the back-propagation neural network(BPNN)-based sub-pixel mapping method. Different from the original BPNN method that uses a single observed coarse spatial resolution image, the new method integrates multiple coarse spatial resolution images that are shifted from each other to determine the probability of a sub-pixel belonging to each class. The probabilities and land cover fractions are then used to allocate classes for sub-pixels. The proposed method can decrease the uncertainty and errors in BPNN-based sub-pixel mapping. Experimental results show that with both visual and quantitative evaluation, the proposed method can obtain more accurate sub-pixel mapping results.