基于正样本-背景数据的校正混淆矩阵

A calibrated confusion matrix based on positive and background data

提出一种基于正样本和无标记背景数据的混淆矩阵校正方法,并以WorldView-3和Landsat8卫星影像不透水面提取为例,验证其在遥感影像一类分类精度评价中的有效性.实验结果表明,基于接受者操作特征曲线估算的校正常数c在WorldView-3影像中范围为0.3017~0.3103,在Landsat8影像中范围为0.2895~0.3132,分别与对应的真值0.2790和0.3000较为接近;基于正样本-背景数据朴素混淆矩阵的精度指标与传统基于正负二类数据基准混淆矩阵的精度指标值相差较大,而经过c值校正之后的精度指标值与基准指标值基本一致.该结果验证了基于正样本-背景数据的校正混淆矩阵在不依赖负样本的情况下可以有效地对一类分类结果进行评价.

Remote sensing classification has been widely used to monitor the land use land cover change.The sit-uation under which the users are only interested in a single land use type is called as one-class classification.In one-class remote sensing classification,researchers usually use one-class classification methods to reduce the cost of manual labeling,i.e.only samples from the target class(or positive class)are collected without requir-ing samples from other classes(or negative class).However,traditional accuracy assessment methods require both positive and negative data to establish a confusion matrix and calculate accuracy measures.In the report,a calibrated confusion matrix based on the positive and unlabeled background data was proposed,and the two case of mapping impervious surface from a WorldView-3 image and a Landsat8 image were used to evaluate its effec-tiveness in accuracy assessment.The results showed that the calibration constant c obtained from the receiver op-erating characteristic curve falls into the range of 0.3017~0.3103 for WorldView-3 image and 0.2895~0.3132 for Landsat8 image,which are close to the corresponding true values of 0.2790 and 0.3000,respec-tively.There were big differences between the accuracy values obtained from the naive confusion matrix with positive and background data and that from the benchmark values obtained from the traditional confusion matrix with positive and negative data,whereas the accuracy values obtained from the calibrated confusion matrix are very similar to the benchmark values obtained from the traditional confusion matrix with positive and negative da-ta.The results indicated that the proposed calibrated confusion matrix from positive and background data is ef-fective for accuracy assessment of one-class classification,without requiring negative data.