遥感模式分类中的空间统计学应用——以面向对象的遥感影像农田提取为例

Applying Spatial Statistics into Remote Sensing Pattern Recognition:with Case Study of Cropland Extraction Based on GeOBIA

如何有效地从遥感图像中提取所需信息,是遥感图像处理和应用的关键,而尺度选择问题一直是影响遥感信息提取精度的关键问题之一。本文论述了利用空间统计学方法解决遥感影像模式分类中的尺度问题的理论基础。针对面向对象影像分析问题,将影响遥感影像多尺度分割的尺度分割参数概括为空间属性分割参数、光谱属性分割参数和影像对象面积阈值参数,并分别提出了基于统计学的尺度参数估计方法。以SPOT-5影像面向对象农田提取为例,基于变异函数方法进行了尺度优选试验,系列尺度分类试验结果表明基于空间统计学尺度估计得到的尺度分割结果进行分类能得到最高的精度,进而证明了基于空间统计学方法进行面向对象信息提取尺度估计的有效性。该方法是完全数据驱动的方法,基本不需要先验知识参与。不同于以往分割后评价的尺度选择方法会占用大量计算资源且耗费大量时间,本文提出的方法不仅能在一定程度上保证面向对象信息提取的精度,而且在一定程度上也提高了面向对象信息提取的效率和自动化程度。

Information extraction from remote sensing image is the key to remote sensing applications and scale selection is one of the key factors influencing the information extraction accuracy.This paper discusses the theoretical foundation of applying spatial statistical methods to resolve the scale related issues involved in remote sensing pattern classification.Aiming at geo-object-based image analysis（GeOBIA）,scale parameters involved in multi-scale segmentation for GeOBIA are generalized into three ones,and they are spatial parameter,attribute parameter and merging threshold.Further,the preestimation method of the optimal scale parameters is proposed based on spatial statistics.Taking GeOBIA based cropland extraction from SPOT-5 image as an example,this paper proposes a cropland extraction method combining spatial statistics based adaptive scale parameter pre-estimation and object-oriented classification.This paper employs mean-shift segmentation and series Rof object based classification on different scales to verify the validity of this method.Experimental results support the object based cropland extraction method based on the data-driven scale pre-estimation.The cropland extraction result by using the pre-estimated segmentation parameters can guarantee the accuracy of GeOBIA classification and the cropland extraction based on GeOBIA and adaptive scale pre-estimation avoids the time-consuming trialand-error practice and speeds up the object-oriented classification procedure.