10 1 20061 JOURNAL OF REMOTE SENSING Vol. 10, No. 1 Jan., 2006 : 100724619 (2006) 0120039210 1, 2, 1, 3, 1 (1., 100101; 2., 100039; 3., 475001) :,,,,,,,,,,,, :, : ;; : TP751. 1 : A A New Approach to Accuracy A ssessm en t of C la ssif ica tion s of Rem otely Sen sed Da ta ZHENG M ing2guo 1, 2, CA I Q iang2guo 1, Q IN M ing2zhou 3, YUE Tian2xiang 1 (1. Institute of Geographic Sciences and N atural Resources R esearch, CAS, B eijing100101, China; 2. Graduate School of the Chinese A cadem y of Sciences, B eijing100039, China; 3. College of Environm ent and Planning, Henan U niversity, Henan Kaifeng475001, China) Abstract: Accuracy assessment is an indispensable step in the p rocess of classification of remotely sensed data. : 2004211216;: 2005201225 : (40271075), (CXIOG2A04210) : (1971),,, The common method is carried out through confusion matrix established on reference data, which has three deficiencies: the heavy workload, inability to guarantee the comp lete correctness of reference data, the cost of reduction error resulting in the increase of workload. In remotely sensed imagery, the feature vector belonging to one category obeys the normal distribution. Based on this hypothesis and statistic theory, a new method is p roposed established on category distribution. The reference data is unnecessary for p roposed method. For the supervised classification, the workload is extremely little. The key to the p roposed method is that the category population can pass the hypothesis test of a certain distribution, in this case, p roducer s accuracy can be figured out easily. Given the number of the category population, the user s accuracy can be figured out too, and then the overall accuracy can be estimated by user s accuracy and area p roportions of all categories after classification. Finally, the p roposed method in this paper was app lied to image classification for Zhengzhou city as an examp le. The result shows: if the distribution of category population can be given, p roducer s accuracy obtained by common method and p roposed method comp letely conform s in the perspective of statistics. Keywords: accuracy assessment; remote sensing classification; hypothesis testing
40 10 1,,, [ 1 ] [ 2 ],,, [ 3 ],,, ( Confusion matrix) [ 4 ] (Overall accuracy) Kappa, ( Producer s accuracy) (U ser s accuracy) [ 5 ], [ 6 ], [ 4 ],:, [ 7 ] [ 8, 9 ], [ 1012 ],,,,,, 100%,,, [ 2 ], 100% [ 4 ],,,,, ( ), [ 4 ], [ 3 ] ;, [ 6 ],,,,ERDAS IMAGINE, 2 2. 1,, [ 13, 14 ],:,( Spectral class) uu u,,,, [ 15 ],,, u,,,, (Maximum likelihood classification ) 100%,,,,, 1 1 L i,l,a 1, a 2,, a n, ( a 1 ), ( a 2 ),, ( a n ) i,1 i L, X, X,,,,X,, L i (S ) ;L
1 : 41 2. 1. 3 S 1 = S p 1 + S p 2 + + S p n, i 1 Fig. 1The real histogram of category population and the category histogram after classification i i,; (S 1 ) i i, ;L (S 2 ) i S i ( pa),( ua) : pa = S 1 /S (1) ua = S 1 / S (2),S, S 1, S 2, S 2. 1. 1 S 1 S i, i ( a 1, a 2 ), ( a 2, a 3 ),, ( a n, a n + 1 ),a 1, a 2,, a n S, S i ( a i, a i + 1 ) p i, i : 2. 1. 2 pa i = p 1 + p 2 + + p n (3) n, r 1, r 2,, r n, 2. 1. 3 ua 1, ua 2,, ua n, : = r 1 ua 1 + r 2 ua 2 + + r n ua n (4) : ua i = S 1 /S = S pa i /S (5) S,: i, X i p x,x N X, X i, i : S = N X / p x (6), N X i,(6)s, i i p, N, i : S = N / p (7), N,, N, (7) S 2. 2, 2,,,, ;, 3 3. 1, 1100km 2, Landsat25 TM, 200053156,,,,,,
42 10 2 Fig. 2Flowchart of accuracy assessment by the method p roposed in this paper 3. 2 9919% [ 16 ],,, 1998, 4, 5, 3, ( Tasselled cap) ( Principal component),,1 2,,,,, 4, 5, 3,( )( ) ( )121 (, ) 2 (, )3 () /// 12 3. 3 3. 3. 1 Erdas AO I(A rea of interest)(region grow AO I),,, Erdas( Image alarm ),,,, 1,, [ 12, 13 ] 7, 572 3. 3. 2,,,1, 2, 7 1, 2, 7,,3
1 : 43 1 Table 1The num bers of p ixels and polygon s in tra in ing area s 1 2 3 2 1 / / / 8 16 8 3 7 5 7 4 8 5 9 2 4646 4247 493 221 403 1010 1377 2498 381 2878 2420 146 3 Fig. 3Pixls assigned to high density in three classification results,,,, 3. 4 3. 4. 1,, 2
44 10 2 (1, 2, 7) Table 2The process and result of hypothesis testing of popula tion d istr ibution ( band 1, 2, 7) m i p i np i (m i - np i ) 2 / np i 1 2 7-133 11 010059 2713 91752321 133134 56 010098 4513 21513671 134135 108 010213 9817 01865253 135136 226 010406 18815 71453435 136137 316 010679 31512 01001626 137138 437 010994 46119 11352129 138139 564 011277 59311 11431003 139140 635 011436 66711 115534 140141 665 011415 65715 01084166 141142 582 011222 56718 01354459 142143 457 010925 42915 11748555 143144 305 010613 28417 11437471 144145 162 010356 16513 01069379 145146 72 010181 8411 11755847 146147 33 010081 3715 0154391 147148 13 148149 3 17 010047 2116 1100992 149+ 1-56 5657 3 60 010189 8811 81984600322 57 5758 289 010550 25517 41330403689 5859 627 011327 61618 01165946439 5960 1016 012183 101416 01001921028 6061 1151 012449 113810 01147599115 6162 863 011873 87015 01066079197 6263 425 010977 45411 11870933943 6364 136 010347 16115 41027187982 6465 68 010084 39113 21129402747 65+ 11 010015 712 11951786712-56 5657 4 30 010090 4211 314901678 26 5758 51 0100962356 4417 018845885 5859 132 0101721749 7919 331813026 5960 158 010284748 13212 419949569 6061 183 0104353209 20212 118322143 6162 222 0106152027 28518 141251465 6263 333 0108036908 37313 413700013 6364 441 0109705584 45019 012182969 6465 475 0110834632 50313 115997059 6566 544 0111180727 51914 111596365 6667 486 0110665638 49515 011831108 6768 431 0109405162 43619 010813965 6869 399 0107666654 35611 511445809 6970 319 0105777069 26814 915382525 7071 216 0104024121 18619 415104874 7172 115 0102591158 12013 012408972 7273 60 010154232 7116 118960859 7374 27 0100848617 3914 319167314 7475 12 0100431625 2010 312341647 7576 6 0100202937 914 112466834 7677 5 0100143326 616 010652018 6 77+ 1 0100143326 616 010652018 2 = 3119,, 2 161, 2 < 2 0. 005 ( 15 ) ( 2 0. 005 ( 15 ) = 3218), 01005 2 = 42184, 9 (10), 2 > 2 0. 001 ( 9 ) ( 2 0. 001 ( 9 ) = 27187 ), 2 = 96167, 20 (21 ), 2 > 2 0. 001 ( 20 ) ( 2 0. 001 ( 20) = 45131 ), : m i i (133133134) p i i, n4646 (1), np i 2( ),5, 2
1 : 45 2, [ 15 ] : F ( X ),, 2 =(m i - np i ) 2 / np i (8) 2 (Chi2Square) m i i, p i F (X ) i, n 1 139139,21735,, 13918921735,1 13918921735,, 2 6013 1159, 76514 31555 2 2 2, m i np i,,, 2 3. 4. 2,, (3) 3,,2 3 Table 3Fea ture vector and num bers of p ixels cla ssif ied in to h igh den sity urban area 1 2 7 62 31119 140 37011 63 32861 141 30177 60 103334 64 33060 142 24863 65 32344 66 32263 92051 103334 161647 4 3, p x 2p i,, 1 7 1 7 7, 1 (2) 7, 2p i 1, p x 3,3, 4 4,,,, 1 7,1,, ( Spectral class) ( Practical class), (6),N X
46 10 4 Table 4Cacula tion of the num ber of ca tegory popula tion for h igh den sity urban area 1 7 62 0. 146 0. 192 387201 140 0. 390 0. 402 261562 63 0. 194 0. 203 338578 141 0. 342 0. 327 246948 64 0. 208 0. 204 305133 142 0. 268 0. 270 268789 65 0. 239 0. 200 289284 66 0. 213 0. 199 302495,4,,, 3,,,, a,x, n (x ), n, a 3. 4. 3 3,,,,,3, 7,1,2 1, 2x 1, x 2, x 3 011415, 011222, 010925, 013562, ( 3), 013562 2 7,, 1, : 2 012449, 7015042 3. 4. 4, 5 3. 4. 35,,,,, [ 17 ] : U = ( p - P 0 ) / [ P 0 (1 - P 0 ) / n ] 1 /2 (9) U n, p, P 0, 5, 572 (5 ), : [ 18 ], [ 19 ],,,,,,,
1 : 47 5 Table 5The cla ssif ica tion result of test area s 1 2 3 2 1 / / / 1 212 42 0 0 41 0 0 48 0 216 0 13 572 0. 3706 2 166 2 0 61 141 1 73 0 0 13 1 114 572 0. 2902 7 221 11 0 260 58 10 0 0 1 0 0 11 572 0. 386 :42 1,572 42 1 013562, 013706,: 013706, 013562? (9) : U = 0. 3706-0. 3562 [0. 3562 (1-0. 3562) /572 ] 1 /2 = 0. 7192, a0105 1196, U < 1196, U,,,, 2 U = 21502, a 0101 2158,012449 7 U = - 51656,, 015042 4 (1),, ( ),, 1,2,7 12,,,( ), : (1); (2), ; (3) TM,,,, (2) :,,,,,,,,,, (3),,,,,,,, [ 4 ],,,,, (4),,
48 10,,,, ( References) [ 1 ] Gan F P, W ang R S, W ang Y J, et al. The Classification Method for Land U se and Land Cover Based on Remote Sensing Technology [ J ]. Rem ote Sensing for Land & Resource, 1999, 42 (4) : 3944. [,,. [ J ]., 1999, 42 (4) : 3944. ] [ 2 ] Anssen L L F, Vanderwel F J M. Accuracy A ssessment of Satellite2derived Land2cover DataA Review [ J ]. Photogramm etric Engineering and Rem ote Sensing, 1994, 60 (4) : 419426. [ 3 ] Kalkhan M A, Reich R M. A ssessing the Accuracy of Landsat Thmeatic Mapper Classification U sing Double Samp ling [ J ]. International Journal of Rem ote Sensing, 1998, 19 (11) : 2049 2060. [ 4 ] Congalton R G. A Review of A ssessing the Accuracy of Classifications of Remotely Sensed Data [ J ]. Rem ote Sensing of Environm ent, 1991, 37 (1) : 3546. [ 5 ] Hay A M. The Derivation of Global Estimation from a Confusion Matrix [ J ]. International Journal of Rem ote Sensing, 1988, 9 (8) : 13951398. [ 6 ] Congalton R G. A Comparison of Samp ling Schemes U sed in Generating Error Matrices for A ssessing the Accuracy of Map s Generated from Remotely Sensing Data [ J ]. Photogramm etric Engineering and R em ote Sensing, 1988, 54 (5) : 593600. [ 7 ] Chen F, Cheng G, Bao H S, et al. Analysis on Land use Chang and Human D riving Force in U rban Fringe [ J ]. Journal of N atural Resources, 2001, 16 ( 3) : 204209. [,,. [ J ]., 2001, 16 (3) : 204209. ] [ 8 ] L i X, Ye J A. Anthony Gar on Yeh. Accuracy Imp rovement of Land U se Change Detect on U sing Principal Components Analysis: A Case Study in the Pearl R iver Delta [ J ]. Joural of Rem ote Sensing, 1997, 1 (4) : 282289. [,. [ J ]., 1997, 1 (4) : 282 289. ] [ 9 ] Q iao Y L. Study on Remote Sensing Classification Method of Classifying H igh, Medium and Low Yield Crop Lands and Their Form ing Factors in the Loess Plateau Taking D ingxiang Country of Shanxi Province as an Example[ J ]. Journal of Rem ote Sensing, 2002, 6 (1) : 7075. [. [ J ]., 2002, 6 (1) : 7075. ] [ 10 ] Shu H L, Mao Z Y. Knowledge Based Image Classification App roach Supported by GIS [ J ]. Acta Geodaetica et Cartographical S inica, 1997, 26 ( 4) : 328336. [,. GIS / [ J ]., 1997, 26 ( 4) : 328336. ] [ 11 ] L iu Y L, Yan S Y, W ang T, et al. Study on Segmentation Based Classification App roaches for Remotely Sensed Imagery [ J ]. Journal of Rem ote Sensing, 2002, 6 (5) : 358363. [,,. [ J ]., 2002, 6 (5) : 358363. ] [ 12 ] Zhou B, Yang B L. The Research on Land U se Change Detection by U sing D irect Classification of Stacked Multi Temporal TM Images[ J ]. Joural of N atural R esources, 2001, 16 ( 3 ) : 263 267. [,. [ J ]., 2001, 16 ( 3 ) : 263267. ] [ 13 ] Peng W L. The Computer Data Processing of Remote Sensing Data and Geographical Information System [M ]. Beijing: Beijing Normal University Press, 1991. [Θ. [M ]. :, 1991. ] [ 14 ] Molk J G. D igital Processing of Remotely Sensed Images[M ]. Beijing: W eather Press, 1987. [ J G. [M ]. :, 1987. ] [ 15 ] Sheng H F. The Tutorial of Theory of Probability and Mathematical Statistics [M ]. Beijing: H igher Education Press, 1998. [. [M ]. :, 1998. ] [ 16 ] L iu J G. Pixel B lock Intensing Modulation Adding Spatial Detail to TM Band6 Thermal Imargy [ J ]. Rem ote Sensing, 1998, 19 (3) : 24772491. International Journal of [ 17 ] Geng X L, Xie Z R. App lied Statistic [M ]. Beijing: Science Press, 2002. [,. [M ]. :, 2002. ] [ 18 ] ShiW Z. Theory and Methods for Handling Errors in SpatialData [M ]. Beijing: Science Press, 2000. [. [M ]. :, 2000. ] [ 19 ] N i J X. App lied Statistic [ M ]. Beijing: Renm in University Press, 1994. [. [M ]. :, 1994. ]