CAJ | 학술논문

为了进一步提高农作物遥感识别精度，充分利用高分辨率遥感影像上不同地物之间的邻域空间关系，提出农作物遥感识别偏差修正的地统计学方法。该方法综合考虑目标地物的光谱特征与空间信息，以类别隶属度偏差为研究对象，首先利用类别指示向量和类别后验概率向量之间的差异实现目标地物的类别隶属度偏差量化，然后对训练样本的类别隶属度偏差进行变异函数建模，并采用带局部均值的简单克里格插值方法预测总体类别隶属度偏差，之后用总体偏差的预测值对光谱分类所得的类别后验概率进行修正，重新确定识别结果，实现农作物遥感识别的偏差修正。以安徽省南部的一景 SPOT-5影像覆盖范围为研究区，选择2块典型区域分别作为试验区和验证区，以一季稻和晚稻为目标农作物，以支持向量机作为光谱分类的分类器，建立了水稻遥感识别的偏差修正流程；采用地面实测数据对修正效果进行评估，并与最大似然分类、模糊分类和支持向量机分类的结果进行比较。试验结果表明，该方法的总体分类精度能够达到90%以上，与传统分类方法相比，总体精度提高了近14%；且该方法能够大幅提高一季稻和晚稻的生产者精度和用户精度，有效改善了研究区的水稻识别结果，可以为中国南方复杂种植条件下的水稻识别提供参考。
위료진일보제고농작물요감식별정도，충분이용고분변솔요감영상상불동지물지간적린역공간관계，제출농작물요감식별편차수정적지통계학방법。해방법종합고필목표지물적광보특정여공간신식，이유별대속도편차위연구대상，수선이용유별지시향량화유별후험개솔향량지간적차이실현목표지물적유별대속도편차양화，연후대훈련양본적유별대속도편차진행변이함수건모，병채용대국부균치적간단극리격삽치방법예측총체유별대속도편차，지후용총체편차적예측치대광보분류소득적유별후험개솔진행수정，중신학정식별결과，실현농작물요감식별적편차수정。이안휘성남부적일경 SPOT-5영상복개범위위연구구，선택2괴전형구역분별작위시험구화험증구，이일계도화만도위목표농작물，이지지향량궤작위광보분류적분류기，건립료수도요감식별적편차수정류정；채용지면실측수거대수정효과진행평고，병여최대사연분류、모호분류화지지향량궤분류적결과진행비교。시험결과표명，해방법적총체분류정도능구체도90%이상，여전통분류방법상비，총체정도제고료근14%；차해방법능구대폭제고일계도화만도적생산자정도화용호정도，유효개선료연구구적수도식별결과，가이위중국남방복잡충식조건하적수도식별제공삼고。
Crop identification with high resolution satellite imagery relates to four key factors: 1) crop phenologies, which lead to the similarity of plant reflectance of different crops;2) the high resolution, which leads to field-to-field variability of plant reflectance of the same crops; 3) performances of various classifiers, which directly restrict crop identification accuracy; 4) feature variables, which reflect spatial and spectral variability within fields. Considering restrictions by these factors, a geostatistical approach to bias calibration of crop identification is proposed, in order to make full use of spatial structure information in high resolution satellite imagery to further improve classification accuracy of some stable crops like single-cropping rice and late rice. <br> Taking into account spectral characteristics and the spatial structure information of crops, the proposal method is based on the techniques of variogram and Kriging algorithms from geostatistics. First, the differences between indicator vectors and posterior category probability vectors for the training samples were calculated to quantify the biases of category memberships before and after the spectral classification. Then, the generated biases were regarded as regionalized variables, and some kind of experimental variogram models were selected to biases modeling, and the method of simple Kriging with local means was used to predict the biases for all pixels. Finally, the predicted biases were added to the posterior category probabilities derived from the initial spectral classification to obtain the calibration results, and the process of bias calibration of crop identification was then found. <br> A SPOT 5 image acquired in September 2012 with four spectral bands and 10-m pixel size covering intensively cropped areas in south Anhui province was used for crop identification. Two subset images that covered Congyang county and Guichi county with the same area of 100km2 were also generated from the original image as the study area and verification area, respectively. A support vector machine classifier was used to get the spectral classification and posterior category probabilities. Ground truth data were collected and used to evaluate the calibration effects, and the transformations between category memberships before and after calibration were analyzed, with respect to the two major crops. Comparing with the direct spectral classification results generated from methods of maximum likelihood classification, fuzzy classification and support vector machine classification, the overall accuracy of the calibration method increased by nearly 14%, and was always able to achieve above 90%. Moreover, there were some substantial increases in the producer’s accuracy and the user’s accuracy of single-cropping rice and late rice, with the precision of increase more than 30%, effectively improved the identification accuracy of rice in the research region. Therefore, it is illustrated that the proposal method overcomes the limitations due to spectrum characteristics and a similar operation can usually be implemented for crop identification. <br> Based on the direct spectral classification, the proposal method focused on biases of the category memberships of several major crops, and considered the structural and the random characteristics of the posterior category probabilities due to the spatial distribution of categories in the local regions, and thus is independent of the limitation of spectral similarities of some crops. With regard to the operation processes, further improvement upon the calculation of parameters of variogram models will be a future concern, and the optimal sampling strategy will be studied more, considering the spatial distribution of the samples.