自适应的IDW插值方法及其在气温场中的应用
作者简介:段平(1984- ),男,湖北监利人,博士研究生,主要从事空间数据插值方法研究。
收稿日期: 2013-12-24
要求修回日期: 2014-03-18
网络出版日期: 2014-08-10
基金资助
国家自然科学基金项目(41271383)
江苏省普通高校研究生科研创新计划资助项目(CXLX13-376)
南京师范大学研究生科研创新计划资助项目(CXLX13-376)
Adaptive IDW interpolation method and its application in the temperature field
Received date: 2013-12-24
Request revised date: 2014-03-18
Online published: 2014-08-10
Copyright
反距离权重(Inverse Distance Weighting,IDW)插值通常采用距离搜索策略选择插值参考点,当采样点集分布不均匀时,距离搜索策略使得参考点聚集一侧影响插值精度。自然邻近关系具有良好的自适应分布特性,可有效地解决参考点分布不均匀问题。结合自然邻近关系,提出自适应的反距离权重(Adaptive-IDW,AIDW)插值方法。首先对采样数据构建初始Delaunay三角网,然后采用逐点插入法,将待插值点插入初始Delaunay三角网中,局部调整得到新的Delaunay三角网,以待插值点的一阶邻近点作为IDW插值的参考点,使参考点自适应均匀地分布在待插值点周围,再进行IDW插值计算。利用AIDW插值方法对Franke函数、全国气温观测数据进行插值实验,结果表明此方法具有较高的精度,且减少了“牛眼”现象。
段平 , 盛业华 , 李佳 , 吕海洋 , 张思阳 . 自适应的IDW插值方法及其在气温场中的应用[J]. 地理研究, 2014 , 33(8) : 1417 -1426 . DOI: 10.11821/dlyj201408003
The Inverse Distance Weighting (IDW) interpolation has the advantage of simpleness, convenience for calculation, and high compatibility with Tobler's first law. It is widely used in construction of DEM, weather analysis, hydrological analysis, and so on. Distance search strategy is usually adopted by the IDW interpolation to select referent points. However, referent points gathering in one side may cause the loss of interpolation accuracy when sampling points are unevenly distributed. The natural adjacency spatial relationship with good adaptive characteristics about choosing referent points can effectively solve the problem of reference points’ uneven distribution. Based on this, the adaptive inverse distance weighting (AIDW) interpolation method was proposed in this paper. Firstly, the initial Delaunay triangulation was built with the sampling data points. Secondly, interpolative points were inserted one by one, the purpose of which was making the referent points evenly distributed around the interpolative points by taking the first-order neighboring of interpolative points as referent points. At last, IDW was interpolated. In step two, when each point was inserted, the Delaunay triangulation should be reconstructed, which elapseded time a lot. To solve this problem, the patial grid index was built in order to raise the speed of Delaunay triangulation's reconstruction. Compared with ordinary IDW, there was no need to assign the number of referent points or search radius in the AIDW, because the referent points could be adaptively selected with the natural adjacency spatial relationship of interpolative points. Especially when referent points were too intensive, the problem of superfluous points being inserted could be avoided. Two experiments were conducted in this paper, which were the theoretical surface reconstruction of the Franke and the national surface air temperature field reconstruction respectively. The results were compared with IDW interpolation methods with different search strategies in ArcGIS 10.1, which verify that the proposed method have higher accuracy. Meanwhile, the results of the proposed method show that the 'buphthalmos' phenomenon is reduced. All the outcomes indicate that the proposed method can be applied in the interpolation of geographical phenomenon as a new method.
Key words: IDW; Delaunay; natural neighbor; interpolation; temperature
Fig.1 The search strategy图1 搜索策略 |
Fig.2 Flow chart of the AIDW图2 AIDW流程图 |
Fig. 3 The regular grid index图3 规则格网索引 |
Fig. 4 Selections of the first order natural neighbors图4 自适应选取一阶自然邻近点 |
Fig.5 Franke surface<![CDATA[(2)]]> 图5 Franke曲面 |
Tab. 1 Interpolation error statistics of 500 sampling points and 1000 interpolation points表1 采样点为500、插值点为1000的插值误差统计 |
| 插值方法 | εmax | εmin | εme | εrmse | 时间(s) |
|---|---|---|---|---|---|
| IDW/点数固定 | 1.0165E-01 | -7.6662E-02 | 1.1487E-03 | 1.8337E-02 | 9.1131E-02 |
| IDW/半径固定 | 1.1327E-01 | -9.5031E-02 | - | - | 5.9191E-02 |
| NNI | 8.5239E-02 | -1.0396E-01 | -1.1884E-03 | 1.1155E-02 | 1.3415E+00 |
| AIDW | 6.7655E-02 | -6.2278E-02 | 1.0564E-04 | 1.2819E-02 | 1.4808E-01 |
Tab. 2 Interpolation error statistics of 1000 sampling points and 1000 interpolation points表2 采样点为1000、插值点为1000的插值误差统计 |
| 插值方法 | εmax | εmin | εme | εrmse | 时间(s) |
|---|---|---|---|---|---|
| IDW/点数固定 | 6.7154E-02 | -5.2808E-02 | 8.7642E-04 | 1.1987E-02 | 1.2359E-01 |
| IDW/半径固定 | 7.2202E-02 | -6.4808E-02 | - | - | 9.8902E-02 |
| NNI | 6.4723E-02 | -8.3556E-02 | -4.6503E-04 | 6.8004E-03 | 2.8475E+00 |
| AIDW | 3.8223E-02 | -4.1399E-02 | 3.7022E-04 | 8.2223E-03 | 2.9629E-01 |
Tab. 3 Interpolation error statistics of 2000 sampling points and 1000 interpolation points表3 采样点为2000、插值点为1000的插值误差统计 |
| 插值方法 | εmax | εmin | εme | εrmse | 时间(s) |
|---|---|---|---|---|---|
| IDW/点数固定 | 4.7040E-02 | -3.9979E-02 | 1.4654E-04 | 8.1831E-03 | 1.9661E-01 |
| IDW/半径固定 | 4.4330E-02 | -3.9197E-02 | 1.0089E-04 | 7.8569E-03 | 1.0591E-01 |
| NNI | 5.2840E-02 | -6.7897E-02 | -2.2281E-04 | 4.3300E-03 | 5.9657E+00 |
| AIDW | 2.5195E-02 | -2.8181E-02 | 1.0064E-04 | 5.2229E-03 | 1.0079E-01 |
Tab. 4 Interpolation error statistics of 5000 sampling points and 1000 interpolation points表4 采样点为5000、插值点为1000的插值误差统计 |
| 插值方法 | εmax | εmin | εme | εrmse | 时间(s) |
|---|---|---|---|---|---|
| IDW/点数固定 | 2.9507E-02 | -2.4713E-02 | 9.9800E-05 | 5.0458E-03 | 4.5276E-01 |
| IDW/半径固定 | 3.1048E-02 | -2.3046E-02 | 8.8139E-05 | 4.8384E-03 | 3.3553E-01 |
| NNI | 3.8267E-02 | -3.5618E-02 | -9.7021E-05 | 2.4240E-03 | 1.5587E+01 |
| AIDW | 1.7637E-02 | -1.6134E-02 | -1.4573E-07 | 3.0991E-03 | 2.0731E-01 |
Tab.5 Interpolation error statistics of 10000 sampling points and 1000 interpolation points表5 采样点为10000、插值点为1000的插值误差统计 |
| 插值方法 | εmax | εmin | εme | εrmse | 时间(s) |
|---|---|---|---|---|---|
| IDW/点数固定 | 2.0089E-02 | -2.0537E-02 | 4.6727E-05 | 3.5117E-03 | 8.1887E-01 |
| IDW/半径固定 | 2.4847E-02 | -1.5791E-02 | 6.8011E-05 | 3.4972E-03 | 1.9138E-01 |
| NNI | 3.0395E-02 | -7.8220E-03 | -2.5906E-05 | 1.2156E-03 | 3.3864E+01 |
| AIDW | 1.1890E-02 | -1.1242E-02 | 7.7898E-06 | 2.0983E-03 | 1.3803E-01 |
Tab. 6 Interpolation error statistics of 20000 sampling points and 1000 interpolation points表6 采样点为20000、插值点为1000的插值误差统计 |
| 插值方法 | εmax | εmin | εme | εrmse | 时间(s) |
|---|---|---|---|---|---|
| IDW/点数固定 | 1.2244E-02 | -1.1579E-02 | 3.7575E-05 | 2.3602E-03 | 9.3095E-01 |
| IDW/半径固定 | 2.1479E-02 | -1.1672E-02 | 9.4939E-05 | 2.6900E-03 | 3.2508E-01 |
| NNI | 1.3414E-02 | -4.2618E-02 | -5.1863E-05 | 1.7519E-03 | 7.0862E+01 |
| AIDW | 8.8736E-03 | -7.3645E-03 | 1.5984E-05 | 1.4228E-03 | 2.1341E-01 |
上,AIDW方法只有20000个参考点的情况下优于NNI。从时间上分析,AIDW明显优于NNI插值方法。在插值点数分别为500和1000时,AIDW方法不及半径固定的IDW;但是当点数大于1000以后,AIDW方法的效率明显都优于其他两种IDW插值方法。其主要原因是三角网构建中使用了格网索引,虽然在数据量较小时与不建立格网索引算法区别不大,但随着数据量的增加,格网索引在查找待插值点一阶自然邻近时发挥了优势,大大减少了计算时间。NNI插值算法时间效率很低,与以上三种IDW方法均不在一个数量级上。Fig.6 Spatial distribution of the data图6 数据空间分布 |
、
、
、
作为误差统计指标,其中交叉验证法采用pick-one-out[16],即每次假设一个采样点未知,将余下采样点通过某种插值方法进行估算,计算估算值与真实值误差,对每一个采样点进行相同的假设和插值估算。分析AIDW与ArcGis 10.1中各种搜索策略下IDW方法的误差统计。各插值方法的交叉验证误差统计如表7所示。Tab.7 The error statistics of interpolation methods (oC)表7 各种插值方法的误差统计(oC) |
| 方法 | εmax | εmin | εme | εrmse |
|---|---|---|---|---|
| IDW-Default | 3.8939E+00 | -3.7626E+00 | 2.5680E-02 | 7.9266E-01 |
| IDW-4/2 | 4.0916E+00 | -3.7738E+00 | 1.5773E-02 | 7.6070E-01 |
| IDW-8/1 | 3.9136E+00 | -3.7762E+00 | 1.4010E-02 | 7.5865E-01 |
| IDW-8/2 | 3.7270E+00 | -3.6215E+00 | 1.5546E-02 | 7.6433E-01 |
| NNI | 4.2397E+00 | -3.5932E+00 | -7.3869E-03 | 7.4926E-01 |
| AIDW | 3.8492E+00 | -3.8634E+00 | 9.3323E-03 | 7.5376E-01 |
不及AIDW方法。NNI插值方法中需要存储Voronoi图,存储结构复杂且计算耗时,在Franke函数模拟实验中已表明NNI插值算法比AIDW插值方法耗时许多,原因是NNI方法插值点需要重新划分权重面积,计算多个多边形的面积,影响了实际应用。图7显示了以上各种方法的插值结果,其中ArcGis 10.1的各种搜索策略IDW插值方法出现“牛眼”现象较为严重,而AIDW方法较好地克服了这种现象。Fig. 7 Comparison of interpolation results图7 各种插值方法比较 |
The authors have declared that no competing interests exist.
| [1] |
[
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
[
|
| [8] |
[
|
| [9] |
[
|
| [10] |
[
|
| [11] |
[
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
[
|
| [16] |
|
/
| 〈 |
|
〉 |