信息素养2.0的内容模块研究

Size: px

Start display at page:

Download "信息素养2.0的内容模块研究"

宏裤阎
5 years ago
Views:

1 遥感数据的时空相关性及在数据预处理中应用报告人 : 陈晋北京师范大学地表过程与资源生态国家重点实验室北京师范大学减灾与应急管理研究院 2012 年 12 月 10 日

2 目录遥感数据的时空相关性条带插补厚云去除山体阴影信息恢复 MODIS/Landsat 数据融合

3 1 遥感数据的时空相关性遥感数据目前已经被地理学生态学等诸多领域作为大尺度不同于普通的数字图像处理研究的重要数据源然而各种因素使得原始遥感数据不能遥感数据的处理过程中需要考虑遥感数据的时空相关性满足大部分研究的需求杜绝将数据孤立地处理因此学者开发了大量数据处理方法数据预处理全球变化生态系统监测

4 1 遥感数据的时空相关性空间相关性 Rs All attribute values on a geographic surface are related to each other but closer values are more strongly related than more distant ones. Tobler s First Law 相关函数距离因子类别因子

5 1 遥感数据的时空相关性时间相关性 Rt 地物的部分属性会随着时间的推移发生周期或者非周期性的变化这些连续的变化便形成了地物的时间相关性间隔时间越短地物属性应该越接近 t 1 t 2 t 3 相关函数距离因子类别因子

6 1 遥感数据的时空相关性空间相关性的时间不变效应在自然情况下若地物类型不变地物的空间相关性应该是不随时间的推移而发生变化

7 1 遥感数据的时空相关性混合像元分解由于地物覆盖类型的复杂性和传感器分辨率的限制混合像元在遥感数据中十分普遍混合像元分解即是获得混合像元中各地物类型比例的一种重要方法端元比例 R mix i f i R i 混合光谱端元光谱

8 1 遥感数据的时空相关性数据缺失问题由于多方面因素的限制遥感数据的缺失问题是遥感数据预处理中面临的最主要的问题数据缺失已有时相数据缺失特定时相数据缺失条带 (ETM+) 厚云山体阴影回访周期

9 1 遥感数据的时空相关性对于数据缺失问题其根本的解决方法是充分利用遥感数据的时空相关性根据已有数据进行插值目标像元相似像元时间相关空间相关相似像元

10 1 遥感数据的时空相关性寻找相似像元确定权重确定插值函数时间空间特征时间空间特征目标像元相似像元

11 2 条带插补 Both Landsats 5 and Landsat 7 are still functioning. Landsats 5 has substantially exceeded its planned design lives On May the scan-line corrector (SLC) for the ETM+ sensor on board Landsat 7 failed permanently. Landsat 5 TM Landsat 7 ETM+

12 2 条带插补 Gap Filling: existing methods linear histogram-matching method (USGS/NASA 2004): have difficulty with heterogeneous landscapes multi-scale segmentation approach (Maxwell et al. 2007): lower accuracy at the pixel level especially for the narrow or small objects Geo-statistics based methods (Zhang et al. 2007; Pringle et al. 2009) using kriging or co-kriging techniques: predict the reflectance not well at the pixel-level and very computationally intensive

13 2 条带插补 For Our Method Assumption: Neighboring pixels in close proximity to SLC-off gaps share similar spectral characteristics and temporal patterns of changes with the missing pixels if they belong to the same land cover type. It is logical to make use of the information of the same-class neighboring pixels to restore spectral values of missing pixels Called as the Neighborhood Similar Pixel Interpolator (NSPI) Two data sources that can be used : (1) TM image or SLC-on ETM+ image and (2) SLC-off images

14 2 条带插补 Using a single TM or SLC-on ETM+ image Target un-scanned pixel (x y) at t 2 Input image at t 1 Search common pixel Search similar pixel Window size+2 No N M No Max window? Yes N>0 Yes Compute weight W Yes No Prediction1 Prediction2 Compute change from common pixel Compute weight T Final prediction Final prediction

15 2 条带插补 Step 1 Step 2 Step 3 Step 4 Selection of neighboring similar pixels RMSD i n b1 ( L( x i yi t1 b) L( x y t b)) n 1 2 Target pixel Similar pixel Common pixel Gap in target image Threshold: RMSD [ n ( b) 2 / m] i / b1 initial moving window size: IWS ( 1)/ 2*2 1 M M is the required sample size maximum window size n

16 2 条带插补 Step 1 Step 2 Step 3 Step 4 Calculation of weight for similar pixels This is determined by the location of the similar pixel and the spectral similarity between the similar pixel and target pixel D j (x j x) 2 (y j y) 2 CD j RMSD j D j W j (1/ CD j ) / N j1 (1/ CD j )

17 条带插补 2 Step 1 Step 2 Step 3 Step 4 Calculation of value of the target pixel two methods to predict the value of the target pixel : first prediction: the weighted average of all the similar pixels: Second prediction: the sum of value at t1 and the change from t1 to t2 ) ( ) ( N j j j j b t y x L W b t y x L )) ( ) ( ( ) ( ) ( N j j j j j j b t y x L b t y x L W b t y x L b t y x L

18 条带插补 2 Step 1 Step 2 Step 3 Step 4 Combine the two predictions with weights: N j n b j j n b t y x L b t y x L N R / ] )) ( ) ( ( [ 1 N j n b j j j j n b t y x L b t y x L N R / ] )) ( ) ( ( [ 1 ) 1/ )/(1/ 1/ ( 2 R 1 R R T i i ) ( ) ( ) ( b t y x L T b t y x L T b t y x L Target pixel Gap in target image Similar pixel Common pixel

19 2 条带插补 All input images are sorted based upon acquisition date. the maximum window size is increased from 17 to 31 Target un-scanned pixel (x y) at t 2 The ith input image Correspondin g input pixel scanned? Yes Target pixel Similar pixel Common pixel Gap in target image i+1 No Predict the value of un-scanned pixel (x y) according to the method of using a single input image Gap in input image

20 2 条带插补 Test results Simulated SLC-off images : test single input image TM 5/25/2008 TM 6/10/2008 Simulated 6/10/2008 TM 2/8/2010 TM 4/29/2010 Simulated 4/29/2010

21 RMSE of NIR 2 条带插补 Optimization of sample size M: Minimum number of similar pixels (M) recommend 20 as an appropriate value of M

22 2 条带插补 Filled results of Simulated SLC-off image 6/10/2008 filled by USGS 6/10/2008 filled by NSPI True image 4/29/2010 filled by USGS 4/29/2010 filled by NSPI True image

23 2 条带插补 Scatter plot of filled image v.s. true image 6/10/2008 4/29/2010

24 2 条带插补 Test results Simulated SLC-off images : test single input image

25 2 条带插补 Test results Simulated SLC-off images : test single input image Simulated 2/8/2010 Simulated 1/23/2010 Simulated 4/29/2010

26 2 条带插补 Test results Simulated SLC-off images : test single input image USGS NSPI 4/29/2010

27 2 条带插补 Test results: True SLC-off images a b c (a) acquired at February ; (b) acquired at June ; (c) acquired at September

28 2 条带插补 Test results: True SLC-off images Using single input USGS NSPI Using multiple inputs USGS NSPI

29 2 条带插补 Test results: True SLC-off images Land cover classification accuracy assessment (James E. Vogelmann USGS) For gap areas overall classification accuracies were 90.8% and 92.7% for gap-filled versus reference data sets respectively

30 2 条带插补 Compared with the existing methods the NSPI can restore the value of un-scanned pixels very accurately especially for heterogeneous landscapes and when there is a longer time interval between the input image and target image; The major improvements of NSPI : better use of useful and relevant information of the scanned pixels; ensure the spatial continuity of the filled results. Potential limitations: require the availability of one or more ancillary TM or ETM+ image(s) unclear how much changing land cover will impact final results

31 3 厚云去除云污染是常见现象影响地表特征的提取 ; 主要 2 类 : 薄云厚云薄云厚云

32 3 厚云去除常用的云处理方法 : 1) 薄云处理 : 恢复云下弱信息比如缨帽变换 FFT 数据融合直方图匹配等方法 ; 2) 厚云处理 : 最常见的为替换法其缺点为 : 反演反射率的误差大视觉上有斑块状

33 3 厚云去除常用的云处理方法 : 1) 薄云处理 : 恢复云下弱信息比如缨帽变换 FFT 发展一种基于多源遥感数据的厚云处理方法得到较为准数据融合直方图匹配等方法 ; 2) 厚云处理 : 最常见的为替换法其缺点为 : 反演反射确的反射率产品率的误差大视觉上有斑块状

34 3 厚云去除为检验方法在 Landsat 影像上模拟厚云 : landsat

35 3 厚云去除辅助数据 : 如果仅有一幅时间接近的清晰 Landsat 影像可用 landsat

36 3 厚云去除云 t1 t2 L( x y t ) L( x y t1) L( x 2 y ) 假设空间临近的同类地物的变化量相同 : L( x y) 可以由云周围的非云像元提供

37 3 厚云去除流程图 t1 时刻 Landsat 无云影像 t2 时刻 Landsat 有云影像步骤 1 自动非监督分类标记云污染区域步骤 2 在缓冲区中迭代搜寻各类地物像元步骤 3 计算各类地物的反射率变化量步骤 4 反演云污染像元的反射率

38 3 厚云去除 Step 1 Step 2 Step 3 Step 4 自动分类结果 dense vegetation thin vegetation high-reflectance bareness low-reflectance bareness water

39 3 厚云去除 Step 1 Step 2 Step 3 Step 4 缓冲区迭代搜寻各类地物 1) 在云周围生成 1 个像元宽度的缓冲区在该缓冲区中搜寻各类地物并标记 ; 2) 如果所有地物类型的像元都找到 ( 而且像元数目大于某一阈值 ) 则搜寻停止 ; 3) 如果还有地物类别没有找到则生成 2 个像元宽的缓冲区继续搜寻直到找到所有地物 1 像元宽的缓冲区对于本例搜寻到第 5 层缓冲区时找到所有 5 类地物的像元 ( 大于 50 个 )

40 厚云去除 3 Step 1 Step 2 Step 3 Step 4 计算各类地物的反射率变化量 m t y x L t y x L L m j j j j j i )] ( ) ( [ 其中 j 是指上一步骤中搜寻到的 i 类地物 m 个像元中的第 j 个像元各类地物各波段的变化量 : 值 /10000

41 厚云去除 3 Step 1 Step 2 Step 3 Step 4 反演云像元的反射率 1) 云边缘的区域 : 即可直接由 t2 时刻相邻同类地物的反射率得到 L1(xyt2) 也可由 t1 时刻的反射率加上变化量得到 L2(xyt2) 用权重将二者结合得到反演结果 : 2) 云中区域 : 由 t1 时刻的反射率加上步骤 3 得到的变化量得到 ) ( w ) ( w ) ( t y x L t y x L t y x L S S S w S S S w L k t y x L t y x L ) ( ) ( 1 2 其中 S1S2 分别表示在目标像元为中心 17*17 窗口内和目标像元同类的地物无云区域面积和被云污染的面积

42 3 厚云去除 Results True image Retrieved image Use 8/12 TM

43 3 厚云去除 Results Green band Red band NIR band

44 3 厚云去除发展的新方法能较为精确地去除影像中的厚云污染反演的的反射率误差很小而且云处理后的影像视觉效果很好没有斑块效果 ; 该方法自动简单快捷可以运用于海量遥感数据云处理

45 4 山体阴影信息恢复 Mountainous shadows occur when objects totally or partially occlude direct light from a source of illumination cause great difficulty in land cover interpretation and classification.

46 4 山体阴影信息恢复 Topographic correction methods can partly alleviate the impact of shadows. But they have tow limitations: 1) little effect on areas with very low incidence angles and those completely with no direct solar illumination (cast shadow); 2) Complete DEM data with adequate spatial resolution and elevation accuracy. Our research is to restore shadow spectral information:

47 4 Shadow Detection 山体阴影信息恢复 Step 1 Step 2 Step 3 Step 4 Step 5 TC transformation Brightness band Brightness of all segmented objectsaverage brightness of all pixels within the object Automatic thresholding method Brightness<threshold shadow object Brightness>threshold non-shadow object

within a certain range of shapes and sizes: Detect shadow

48 4 山体阴影信息恢复 Step 1 Step 2 Step 3 Step 4 Step 5 Image Segmentation: Shadows in mountainous terrain often fall within a certain range of shapes and sizes: Detect shadow at object scale rather than at pixel scale; Image segmentation.

49 4 Shadow Detection Validation 山体阴影信息恢复 Step 1 Step 2 Step 3 Step 4 Step 5 Detected shadow Hillshade derived from DEM Omission error: 2.5%; Commission error: 8.3%; Overall accuracy: 90%

50 4 山体阴影信息恢复 Search for Non-shaded Similar Pixels Continuum Removal (CR) to separate a spectrum into two parts: brightness information (CI) spectral shape information (CR) Step 1 Step 2 Step 3 Step 4 Step 5 CR: divide original spectrum by continuum line CR OI CI

51 4 山体阴影信息恢复 Search for Non-shaded Similar Pixels Step 1 Step 2 Step 3 Step 4 Step 5 Create a buffer with width of two pixels around a shadow object; For a specific target pixel search for N spectrally nearest pixels within the buffer area as the non-shaded similar pixels; If the number of found similar pixels < N repeatedly expand the buffer by two pixels to continue researching until N is met.

52 4 山体阴影信息恢复 Search for Non-shaded Similar Pixels Similarity Condition: RMSD CR (x i y i b): CR value of ith non-shaded pixel located at (x i y i ) in band b; CR (x y b): same as CR (x i y i b) but for the target pixel; n: number of bands; n b1 i CR x y b CRx y b n RMSDi b 2 / m / n b1 Step 1 Step 2 Step 3 Step 4 Step 5 i σ(b): standard deviation of the CR value for the whole test image in band b; m: estimated number of land cover classes. i n 2 Yes similar pixel No not similar pixel

53 4 山体阴影信息恢复 Shadow Information Restoration Step 1 Step 2 Step 3 Step 4 Step 5 Combine the CR of the target shaded pixel and the CI of its non-shaded similar pixels: OI CI CR reb wavgb b OI re(b) : restored DN value of the target shaded pixel at band(b); CR (b) : the continuum removed value of the target shaded pixel at band(b); CI wavg(b) : the weighted average continuum values of non-shaded similar pixels. CI wavg b N j1 W j CI W j : the contribution of similar pixel j. It is related to both the geographic distance between the similar pixel and the target pixel and the spectral similarity between them. j b

4 山体阴影信息恢复 Case study Study area: Shaanxi province China 109 1 24 E 33 57 22 N Forests dominate the mountainous areas in the

54 4 山体阴影信息恢复 Case study Study area: Shaanxi province China E N Forests dominate the mountainous areas in the southern part; Crops urban areas and barren lands cover the northern part of the test area. Landsat 5 TM image: acquired on June

55 4 山体阴影信息恢复 Case study CR method Topographic correction (DEM resolution:90m)

56 4 山体阴影信息恢复 Case study Spectral comparison before and after shadow restoration *The former two curves are presented in TOA reflectance; the latter two are presented in surface reflectance

57 4 Experiments on other areas Origin 山体阴影信息恢复 Topographic correction (DEM:30m) CR method

58 4 山体阴影信息恢复 Validation

59 4 山体阴影信息恢复 Land cover classification Origin Topographic correction CR method OA=75.97% Kappa= OA=77.85% Kappa= OA=86.93% Kappa=0.8235

60 4 山体阴影信息恢复 Texture analysis Window size: 5 5 Indicator: Coefficient of Variance (CV)=standard deviation/mean CR method achieves the smallest CV indicating the most homogeneous image. For mountainous areas dominated by homogeneous vegetation the results clearly show the effect of CR and its advantage over topographic correction.

61 4 山体阴影信息恢复 This research proposes a new method to restore the radiometric information of shaded pixels of mountainous terrain in Landsat TM/ETM+ images without the aid of DEM data. Through simulated spectral assessment and classification experiments on Landsat TM images the proposed method was demonstrated to show improved spectral quality and classification accuracy compared to the original image and the topographically corrected one.

62 4 山体阴影信息恢复 CR method makes full use of the relationship between shaded pixels and their neighboring pixels with the similar spectral shape; CR doesn t require DEM data naturally avoiding the errors originating from this type of data such as resampling and geometric registration; Image segmentation: not only reduce the salt and pepper effect of shaded pixels but also makes it possible to search for non-shaded similar pixels within the buffer area around each shaded object speeding up the search process.

63 5 MODIS/Landsat 数据融合 Temporal Continuous SR: Problem The 16-day revisit cycle of Landsat has limited its use for studying global biophysical processes (16-day 30m) At the same time MODIS scans whole Earth once or twice each day. However the coarse resolution limit its ability in heterogeneous landscapes (daily 250m & 500m) A Fusion Solution To combine the spatial resolution of Landsat with the temporal frequency of coarse-resolution sensors such as MODIS.

64 5 MODIS/Landsat 数据融合 Spatial and Temporal Adaptive Reflectance Fusion Model: Objectives: Fuse high-frequency temporal information from MODIS and Input: high spatial resolution information from Landsat to produce daily Landsat-like surface reflectance MODIS and Landsat surface reflectance pair at tk MODIS surface reflectance M(xiyjt0) at prediction date Predict: MODIS SR Landsat SR 1?? 4 Landsat surface reflectance L(xiyjt0) at prediction date Gao F. J. Masek M. Schwaller and H. Forrest On the Blending of the Landsat and MODIS Surface Reflectance: Predict Daily Landsat Surface Reflectance IEEE Transactions on Geoscience and Remote Sensing vol. 44 no. 8 pp

5 MODIS/Landsat 数据融合 Existing methods Traditional Image fusion methods (HIS PCS) can combine highresolution panchromatic data with multispectral observations acquired simultaneously;

65 5 MODIS/Landsat 数据融合 Existing methods Traditional Image fusion methods (HIS PCS) can combine highresolution panchromatic data with multispectral observations acquired simultaneously; STARFM:Gao(2006) blended Landsat and MODIS data for predicting daily surface reflectance at Landsat spatial resolution failed in heterogeneous areacannot keep the spatial details. Ture image Rebuilt by STARFM

66 5 Theoretical basis MODIS/Landsat 数据融合 Pure coarse-resolution pixel :the difference between the MODIS and Landsat is only caused by the systematic biases Mixed coarse-resolution pixel: based on spectral linear mixing model and the assumption the reflectance change is stable for every land cover type during a short period

MODIS/Landsat 数据融合 5 Theoretical basis Considering the spatial consistency of reflectance change A moving window method is thus used to take full advantage of the information from neighbor pixels

67 MODIS/Landsat 数据融合 5 Theoretical basis Considering the spatial consistency of reflectance change A moving window method is thus used to take full advantage of the information from neighbor pixels especially the information from pure pixels )) ( ) ( ( ) ( ) ( / 2 / 2 / 2 / N i i i p i i i i w w p w w B t y x C B t y x C V W B t y x F B t y x F coarse-resolution pixel central pixel moving window fine-resolution pixel similar pixel

68 5 Process of implementation Flowchart MODIS/Landsat 数据融合

69 5 MODIS/Landsat 数据融合 Step 1 Step 2 Step 3 Step 4 Selection of similar neighbor pixels t m t n

70 5 MODIS/Landsat 数据融合 Step 1 Step 2 Step 3 Step 4 Calculation of weight for similar pixels consider spectral similarity and spatial distance

71 5 MODIS/Landsat 数据融合 Step 1 Step 2 Step 3 Step 4 Calculation of conversion coefficient : apply linear regression model t n t m m=12 V=6.448 R=0.915 P<0.001

72 MODIS/Landsat 数据融合 5 Calculation of reflectance of the central pixel Step 1 Step 2 Step 3 Step 4 ) ( ) ( ) ( (1/ ) ( ) ( 1/ m n k B t y x C B t y x C B t y x C B t y x C T m n k w i w l p l j w j w l k l j w i w l p l j w j w l k l j k ) ( ) ( ) ( 2 / 2 / 2 / 2 / 2 / 2 / B t y x F T B t y x F T B t y x F p w w n n p w w m m p w w

5 MODIS/Landsat 数据融合 Tests with simulated data (Small) TM MODIS ESTARFM STARFM 0.0 0.

73 5 MODIS/Landsat 数据融合 Tests with simulated data (Small) TM MODIS ESTARFM STARFM Relative Errors of Prediction (in percentage) STARFM ESTARFM Object Radius (in meters)

5 MODIS/Landsat 数据融合 Tests with simulated data (Linear) TM MODIS 0.55 0.0 0.25 0.

74 5 MODIS/Landsat 数据融合 Tests with simulated data (Linear) TM MODIS reflectance of the linear object STARFM ESTARFM ESTARFM STARFM 0.30

75 5 MODIS/Landsat 数据融合 Tests with satellite data (Seasonal Changes over Forest)? 5/24/01 7/11/01 8/12/01

76 5 MODIS/Landsat 数据融合 Tests with satellite data (Seasonal Changes over Forest) (a)true image

77 5 MODIS/Landsat 数据融合 Tests with satellite data (Seasonal Changes over Forest)

78 5 MODIS/Landsat 数据融合 Tests with satellite data (Seasonal Changes over Forest) ESTARFM is more accurate.

79 5 MODIS/Landsat 数据融合 Tests with satellite data (Complex Mixture Region)?

80 5 MODIS/Landsat 数据融合 Tests with satellite data (Seasonal Changes over Forest) (a) True image (b) ESTARFM (c) STARFM

81 5 MODIS/Landsat 数据融合 Tests with satellite data (Seasonal Changes over Forest)

82 5 MODIS/Landsat 数据融合 Tests with satellite data (Seasonal Changes over Forest) ESTARFM improved the accuracy more in heterogeneous area.

83 5 MODIS/Landsat 数据融合 Comparing with the original STARFM algorithm it can produce the synthetic fine-resolution reflectance product more accurately especially for heterogeneous landscapes. The improvements includes using a conversion coefficient intersection of similar pixels selection using spectral similarity weighting the change rather than final prediction of each similar pixel. Limitations:cannot accurately predict the shapes changes; Sensors with different band passes may lead to nonlinear relationship; the assumption that reflectance linear change is constant might be not available during a long period; more computational cost than STARFM.

84 The end Thanks for your attention! Thank you!

(baking powder) 1 ( ) ( ) 1 10g g (two level design, D-optimal) 32 1/2 fraction Two Level Fractional Factorial Design D-Optimal D

$(baking powder) 1 ( ) ( ) 1 10g g (two level design, D-optimal) 32 1/2 fraction Two Level Fractional Factorial Design D-Optimal D$ ( ) 4 1 1 1 145 1 110 1 (baking powder) 1 ( ) ( ) 1 10g 1 1 2.5g 1 1 1 1 60 10 (two level design, D-optimal) 32 1/2 fraction Two Level Fractional Factorial Design D-Optimal Design 1. 60 120 2. 3. 40 10