基于HJ-CCD数据和决策树法的干旱半干旱灌区土地利用分类

为了实现干旱半干旱灌区地表信息低成本、高效率的动态监测,利用HJ-CCD数据的多时相和多光谱信息,探讨了平罗县土地利用遥感分类方法。首先建立研究区内典型地物的NDVI时间序列曲线,提取反映该区物候模式的时序特征参数;然后对土壤信息丰富的3月份多光谱影像进行主成分变换,选取第1主成分(PC1)作为光谱特征参数,最后基于分类回归树(classification and regression tree,CART)算法进行决策树监督分类。总体分类精度达到92.26%,Kappa系数为0.91,比最大似然法分类结果精度提高了2.58%。研究表明:构建的NDVI时间序列曲线对研究区内的地类具有较强的代表性,提取的时间维和光谱维的分类参数对各地类均有很好地区分性,CART决策树算法分类结果清晰准确且精度较高。该方法为HJ小卫星在干旱半干旱区等区域的深入应用提供科学依据和实证基础。

Land use classification in arid and semi-arid irrigated area based on HJ-CCD data and decision tree method

Abstract: HJ satellites with the characteristics of high temporal resolution, high spatial resolution and large coverage, can provide the regional land use/cover classification with high accuracy. Pingluo county is in the arid and semi-arid area of northwest China, the climate and human irrigation activities caused complicated land use/cover type and serious soil salinization in the study area. In order to achieve the dynamic monitoring of land surface information with low cost and high precision, a regional land use supervised classification based on the classification and regression tree (CART) algorithm was developed and discussed in Pingluo county using the multi-temporal and multi-spectral information of HJ satellite CCD data. Firstly, high quality HJ-1 CCD data (the interval was about 20 d) were selected, and preprocessed including geometric correction, radiometric calibration and atmospheric correction. The normalized difference vegetation index (NDVI) were calculated and overlapped together. Secondly, the land use types including double crops irrigated land, one crop irrigated land, paddy, sand, saline-alkali soil, forest land, construction land and water were adopted for the two-level classification system, and the training samples were selected to obtain the typical NDVI time-series curve of each land type. Then, the characteristic parameters (including maximum, minimum, range, the difference between the value of the July 29 and the May 10 phases, the difference between the value of the October 10 and the July 29 phases, the mean value of the October 4 to the November 8 phases) which could reflect the phonological pattern in the area were extracted through the analysis of the NDVI time-series curves. Thirdly, the principle component transform of a multi-spectral image in March with ample soil information was performed for improving the separation between the construction land and saline-alkali land when the first principal component (PC1) was chosen for a parameter band for classification. Finally, a CART decision tree classification was implemented by combining the multi-temporal and multi-spectral parameter bands in the area. The decision tree had a total of 102 leaf nodes and could be expressed as "If...Then..." forms. The results showed that the overall precision of this classification method was 92.26%. The Kappa coefficient was 0.91. The accuracy of the paddy field was the highest which reached 98.23%. The accuracies of sand, one crop irrigated land and water were all greater than 90%. Double crops irrigated land, forest land, saline-alkali land, construction land were all greater than 80%. The participation of PC1 had made great contributions in improving the classification accuracy, especially for construction land and saline-alkali land, their accuracy increased 26.34% and 12.14%, respectively. The overall accuracy of CART decision tree classification was increased 2.58% than maximum likelihood classification. The classification accuracy of vegetation improved the most. The results of CART decision tree classification were more accurate and meticulous than maximum likelihood classification (MLC), and it effectively correct the obvious wrong classification results in MLC. The study indicated that the established typical NDVI time series curves based on HJ-CCD data had strong representativeness for each land use type in this region. The extracted time and spectrum dimensional parameters could distinguish between most of the land categories well. The results of CART decision tree classification were more clear and accurate than MLC. The proposed methods in this study had certain feasibility and applicability, and could provide empirical basis for the further application of HJ-1 CCD data in land use/cover and environment monitoring in different scale area, and also give informational and technical supports for the multilevel and comprehensive land resources and environment management by using HJ satellite as the main data source.