基于统计混合模型的遥感影像阴影检测

A shadow detection of remote sensing images based on statistical texture features

为提高阴影检测精度,提出一种新的遥感影像阴影检测方法—将径向基函数神经网络构建的混合模型(称作SMM-RBFNN)应用于遥感影像阴影检测。灰度共生矩阵中的能量、熵、对比度和逆差矩4种统计特征量作为混合模型的输入特征矢量,采用类"期望-最大化"算法(类EM)进行参数估计,训练检测器实现阴影检测。对多幅带有浓厚阴影的遥感影像进行实验,结果表明所提出的方法明显优于传统的高斯背景法和直方图阈值法,能够较好地解决强反射性地物漏检和水体错检问题,能够克服基于阈值思想的检测法需要反复实验选取阈值的缺点。

Shadow detection for high spatial resolution remote sensing images is very critical for image segmentation, feature extraction, image matching, automatic target detection and target location. In order to improve the accuracy of shadow detection, we propose a new shadow detection method based on a statistical mixture model, which combines several radial basis function neural networks. Four statistical features, including energy, entropy, contrast and inverse difference moment, extracted from grey level concurrence matrix are used as the model input features. EM-like algorithm is adopted to estimate the model parameters through optimizing the system cost function. Comparative experiments are performed between the Gaussian background model and the histogram threshold method. Experimental results show that higher detection accuracy of the proposed approach is obtained. The proposed method can solve the problem such as high reflective regions and false alarms in the presence of water, as well as the repeated threshold calculation.