School of Geomatics and Urban Information, Beijing University of Civil Engineering and Architecture, Beijing, China 2

Transcription

1 Examination Metod and Implementation for Field Survey Data of Crop Types Based on Multi-resolution Satellite Images Yang Liu, Mingyi Du, Wenquan Zu, Scool of Geomatics and Urban Information, Beijing University of Civil Engineering and Arcitecture, Beijing, Cina College of Resources Science and Tecnology, Beijing Normal University, Beijing, Cina Abstract. In order to examine te accuracy of large amount of te field survey data wit less accurate, an examination metod based on multi-resolution satellite images was proposed in tis paper. As tere were so large amount of data, stratified random sampling was used to obtain effective samples. Firstly, vegetation index derived from low-resolution satellite images at different times as been adopted as analysis factor. And wave curve carts were drawn wit te vegetation index. From tose carts, te statistics law of wave curves for different crop types was recognized using for crop types classification. Secondly, ig-resolution satellite images were used to correct te area of crop types to get te final classification results. Finally, te accuracy of te field survey data can be calculated by comparing te original survey data wit te final classification results. Moreover, for convenience using, a software as been developed according to te above examination metod. Keywords: Examination metod and implementation, Stratified random sampling, Automatic processing, Crop types, Multi-resolution satellite images Introduction Planting area and yields of crops are te important basis for government's economic policy making. For a long time, te two kinds of traditional metods were used for te statistics of panting area of crops []. One way is te compreensive statistical report coming from statistical and administrative units at various levels step by step []. Anoter way is to sample field survey of reported data. Tis metod is suitable for a large degree on te discrete variables; frequency distribution was igly skewed socio-economic penomena for investigation [3]. But, no matter wicever metod was used, one step can not be omitted, tat is examining te accuracy of large amount data of te field survey. Only in tis way, te data can become te reliable argument for te decision analysis. In recent years, wit te extensive application of remote sensing tecnology, te study of estimation and examination of crop area using remote sensing tecniques as made significant progress, and is steadily moving towards te direction of business

2 [4][5]. Compared wit traditional ground survey and verification, te demand for manpower and costs are significantly reduced, using remote sensing image data [6]. However, from te level of remote sensing tecnology for carrying out, large-scale remote sensing examination of crop area is facing tree problems, namely, precision, efficiency and cost problems [7] [8]. Researc Process Tis article aims to use multi-resolution remote sensing imaging tecnology(two kinds of resolution images: te ig-resolution images, suc as ASTER images, and te low resolution images, suc as MODIS images, Utilizing teir own advantage, combined wit te traditional random, stratified sampling metod [9], te examination sceme for field survey data of crop types was designed, In te sceme, stratified random sampling was used to ensure minimum sample size and stability, and te difference spectrum of low resolution remote sensing images was used to matc te types of corps, and ig-resolution remote sensing images was used to te visual identification of te types and area of corps. We try to solve te confliction of precision and cost, and try to provide metodological guidance for a large-scale examination for field survey data of te types and area [0]. Te flow cart of tecnology route in tis work, see Fig., is expressed as follows: Fig.. Tis sows te flow cart of tecnology route. In order to make te multi-resolution satellite images matced eac oter strictly and be in correspondence wit te field survey data, te procedure of remote sensing data processing sould be done. And stratified random sampling was used to obtain effective samples of field survey data of crop types. Vegetation index derived from low-resolution satellite images at different times as been adopted as analysis factor. After all te data ave processed, te identification and verification sould be done troug visual observation metod. Te statistics law of wave curves of vegetation index at different times was recognized using for crop types classification of te samples. Wile, ig-resolution satellite images were used to correct te classified crop area of samples to get te classification results. Finally, te accuracy of te survey data of crop types can be calculated by comparing te original survey data wit te final classification results using statistical analysis metods.

3 In tis paper, we focus on stratified random sampling in tis work. 3 Stratified Random Sampling Te effectiveness of stratified random sampling can be evaluated along two dimensions. First, it migt reduce te total quantity of measurements required. Second, it can do a reduction of measurements would imply cost savings []. 3. Overview In te sampling metod, stratified sampling is one of te most effective metods. Compared wit simple random sampling, stratified sampling as its several great advantages, suc as te fewer number of samples, iger sampling precision and te lower cost. It is te effective way of large-scale statistical sampling surveys []. So stratified sampling was adopted in te paper. Stratified sampling, also known as type sampling, is one of te most commonly used sampling tecniques in practical work [3][4]. Firstly, overall sample is divided into several strata (groups) in accordance wit certain rules in stratified sampling. Secondly, sampling was done witin eac stratum independently. Te resulting sample is called stratified sample. Accordingly, te sample of eac stratum is independent also [5]. Furtermore, if te sampling metod of eac stratum is simple random sample, te sample is called stratified random sampling. Tus te resulting samples are called stratified random samples. 3. Determination of Stratified Number In practice, te number of strata can not exceed alf of te sample size, because of ensuring tat eac stratum as at least two samples and aving te need to calculate te standard deviation of eac stratum in te metod [6].In tis paper, we stratify te population into six strata. 3.3 Determination of Stratified Boundaries It is usually to determinate te boundaries of strata in accordance wit caracteristics or features of population in te stratification [7]. However, wen te obvious caracteristics or features is not easily recognized, te general way to determinate te boundaries of strata is studying general distribution of some properties in population or studying a relative relationsip of te variable X and te variable Y wic are subfeatures of population by means of some matematical metods. In tis paper, te Optimal Stratification Metod(also known as Accumulative Square Root Metod) is used to determinate te boundaries of strata, wic was proposed by Dalenius and Hodges [8], Firstly, a caracteristic property of population sould be determined, according to wic population is stratified. Secondly,

4 population is sorted by te values of te property of population from small to large. Ten, te sum of te values of te property of population is calculated, and te square root of te sum sould be obtained [9]. Finally, population was divided into several strata by equal division metod of te square root. 3.4 Calculation of Sample Size Tere are two main steps in te calculation of sample size of eac stratum as sow in te below. Computation of te total sample size. Te total sample size calculations generally use te following: n = ry ( ) t L = W S w + N L = In Equation (), n is te total sample size, and is te strata-specific variable wic ranges from to L. L is number of strata. W is te Weigt of stratum, wic can be calculate by means of N (te population for stratum ) divide by te total stratified sample size N. S is te population deviation of stratum, wic is calculated using Equation (). And w is te sampling radio of stratum, wic is calculated using Equation (3). And r is te relative error limit (also as known te confidence interval) for (general set 95%). Wile, Y is te population mean, and t is te percentile of te standard normal distribution (z=0.05 for 95% confidence). S is determined by: N S = ( yi Y ) () N i= Note tat te value of sample i of stratum is marked wit a at as y i, Y is te population mean for stratum. Oter symbols are te same as Equa tion (). And w is determined by: WS (3) w = L W S Were, S is te population standard deviation of stratum, namely te square root of S. Also, oter symbols are te same as Equation () or Equation (). = W S ()

5 Sample allocation in eac stratum. For stratified sampling, te sample size of eac stratum still need to determine after te total sample size is fixed. Wen doing te population estimation, te population variance wic can be estimated is related not only to te variance of eac stratum, but also to te sample size of eac stratum. Tere are a lot of sample allocation metods in practice [0]. You can allocate te samples in accordance wit te radio of between te sample size of eac stratum and te sample size of population, or according to desired overall confidence interval wit minimum total sample size. Te sample allocation of eac stratum was done using te Optimal Sample Allocation, wic was proposed by Neyman []. It is expressed matematically as follows: N S (n n = n sall be as an integer) (4) L N S were, N is te sample size of stratum. Also, oter symbols are te same as Equation () or Equation (3). However, te two problems sould be noted in te Optimal Sample Allocation. Tere are: One is in te case of te sampling ratio f = n / N is very large. Tis situation leads to S (te standard deviation of eac stratum) is relatively large, and n (te sample size of eac stratum) may be larger tan N (te population for stratum ). At tis situation, it required for 00% sampling to te population for stratum. Te oter case is tat n is less tan after te calculation using te Optimal Sample Allocation. Te value of n needs to be set to to reduce te impact of random error on te results and te stability wen te sample size is. = 4 Examination Te process of examination for field survey data of crop types can be mainly divided into te following four steps. Note tat, te experiment of examination itself only use a small part of te test data. 4. Data Te entire test data used in tis paper was sown in Fig..

6 Fig.. Tis sows te experimental data in tis work. From left to rigt is successively vegetation index at different times, ig-resolution satellite images, field survey data and te overlay of te above tree data wic sare te same region in te rectangular box. We ave collected te latest report information of field survey data of crop types of county-level, te multi-resolution satellite images wic can cover te survey region in tis step and prepare for te next step of Sampling. Te field survey data of crop types are Sape-file data, wic sould include te type and area property fields. Hig-resolution satellite data now is ASTER images. But if required, it can be iger resolution images. Vegetation index at different times were derived from MODIS time series data, wic were formed as a multi-band image (namely one time corresponds to one band). And all test data was matced strictly 4. Sampling Fig. 3 sows sampling result comparing wit raw data, and te raw data was covered by te sampling data. So te visible parts of te raw data are not sampled. Fig. 3. Tis sows te sampling result comparing wit raw data. We also do te statistics to te sampling result wic is convenient for customers to find out sampling information to decide weter to re-sampling, sown as te Table below. Table. Statistics of sampling data. Index of Num of Num of Standard Boundaries Strata Population Samples Deviations Means Total is

7 4.3 Crop Types Identification and Areas Correction In tis step, multi-resolution satellite images were used to determine crop types and area of te samples by visual observation metod. So tere are two factors as following: Crop types identification. Wave curve carts were drawn wit te index of bands as X-axis and te value of vegetation index as Y axis, as sown in Fig. 4. Tere are 7 bands in tis work. Fig. 4. Tis sows te wave curve carts of vegetation index. Fig. 5 sow wave curve carts of four different positions. From tose carts, te statistics law of wave curves for different crop types was identified using for crop types classification. Fig. 5. Tis sows te scematic diagram of comparison wit different wave curves. Crop areas verification. It provides a tool to draw polygons and calculate area of tose. By intercomparison between ig-resolution satellite images and samples of te same region, you can get te area of samples by measuring. Te Area of all samples sould be measured, wic as been regarded as ypotetical real value. 4.4 Statistical Analysis It needs to get error situation of te field survey data by doing statistics for result of type identification and area verification of samples. As sown in Table, te number of qualified area, te number of qualified type and te number of qualified bot of eac stratum was obtained by statistics. Also, qualified rate of area, qualified rate of type, and overall qualified rate (area and type are bot qualified) of samples were calculated by anti-deduction.

8 Table. Examination result. Index Of Strata Boundaries Num of Num of Population Samples Num of qualified area Num of qualified type Num of qualified Bot Total is Conclusion is Qualified rate of area is 7.07% Qualified rate of type is 56.63% Overall qualified rate is 36.% In statistics, we can find out te abnormal samples wic are te types is rigt, but te rate of area cange exceeding a certain tresold, as sown in Figure 6.In tis works, te tresold was set to 0%. Tese abnormal samples sould be report to te related departments of doing field survey to redo site verification. Fig. 6. Tis sows te abnormal samples. 5 System Implementation An Examination System for Field Survey Data (ESFSD) as been developed based on te above examination metod. Te system (ESFSD) provides integrated and processoriented functions, suc as data processing, stratified random sampling, sample identification and verification, statistical analysis of samples, and so on. Using te system, you can easily complete te entire process of examination for field survey data of crop types. Te system (ESFSD) adopts C/S (client/server) software system structure, C# as te programming language and SQL Server 000 as back database server. Using ArcEngine as development tools, te application program developed as been closely integrated wit system by UG/Open MenuScript. Figure 7 presents te interface of ESFSD.

9 ESFSD also provides data management statistics information of examination result, and supply better decision supports for re-sampling of samples. Fig. 7. Tis sows te system interface. 6 Conclusion In order to examine te accuracy of large amount of te field survey data, an examination metod based on multi-resolution satellite images was proposed in tis paper. For reducing te number of data and cost in te examination, stratified random sampling was used to obtain effective samples. For determining crop types and area of te samples correctly and quickly, multi-resolution satellite images were used utilizing teir own advantage. Low-resolution satellite images are used to identify te corps type of samples, and ig-resolution satellite images were used to verify te area of samples. Finally, te accuracy of te field survey data can be calculated by comparing te original survey data wit te final examination results by statistics. By te testing of actual data, te examination metod is very effective to solve te confliction of precision and cost. And it also provides metodological guidance for a large-scale examination for field survey data of te types and area. Moreover, for convenience using, te system (ESFSD) as been developed according to te above examination metod. Te results of system testing sow ESFSD is compreensive processing platform wic as realized integrated and process-oriented functions for te examination. And it as strong robustness and stability. Acknowledgments. Te researc performed is supported by Beijing Natural Science Foundation (No: 4009), and is partially funded by Jurisdiction of Beijing Municipality under Grant No. PHR , PHR and PHR Te work was performed under Beijing University of Civil Engineering and Arcitecture (No: ). We wis to tank Guoyin Cai for is careful proofreading of te manuscript.

10 References. B.F. Wu, Q.Z. Li: Crop Acreage Estimation Using Two Individual Sampling Frameworks wit Stratification. Journal of Remote Sensing, 004, Vol. 8, No. 6, pp (004). J.F. Wang, J.Y. Liu, and D.F. Zuan, et al: Spatial Sampling Design for Monitoring te Area of Cultivated Land. International Journal of Remote Sensing, 00, Vol. 3, No., pp (00) 3. L.F. Li, J.F. Wang, and J.Y. Liu: Spatial Sampling Optimized Decision-making of National Land Remote Sensing Investigation. Science in Cina Sec. D Eart Sciences, 004, Vol. 34, No. 0, pp (004) 4. R.M. Narayanan, M.K. Desetty, and S.E. Reicenbac: Effect of Spatial Resolution on Information Content Caracterization in Remote Sensing Imagery based on Classification Accuracy. International Journal of Remote Sensing, 00, Vol. 3, No. 6, pp (00) 5. S.B. Cen: Bulletin of Weat, Corn and Rice Yields Estimate Tecnique wit Remote Sensing. Cinese Science and Tecnology Press, Beijing (993) 6. C.E. Woodcock, A.H. Straler: Te Factor of Scale in Remote Sensing. Remote Sensing of Environment, 987, pp (987) 7. H.Q. Liu: Sampling Metod Wit Remote Sensing for Monitoring of Cultivated Land Canges on Large Scale. Transactions of te Cinese Society of Agricultural Engineering, 00, Vol. 7, No., pp (00) 8. H.S. He, S.J. Ventura, and D.M. Mladenoff: Effects of Spatial Aggregation Approaces on Classified Satellite Imagery. International Journal of Geograpical Information Science, 00, Vol. 6, No., pp (00) 9. G.M. Foody: Status of Land Cover Classification Accuracy Assessment. Remote Sensing of Environment, 00, pp (00) 0. X.F. Jiao, B.J. Yang, and Z.Y. Pei: Paddy Rice Area Estimation Using a Stratified Sampling Metod wit Remote Sensing in Cina. Transactions of te Cinese Society of Agricultural Engineering, 006, Vol., No. 5 pp (006). Y.K. Cen, R. Hu: Te Teoretic Structure and Computerized Realization of Stratified Audit Sampling. Te Teory and Practice of Finance and Economics, 003, Vol. 4, pp (003). S.G. Wang, W.H. Cen, and L.D. Gao: Probability Teory and Matematical Statistic, Science Press, Beijing (000) 3. D.J. Brus, J.J. de Gruijter, J.W. van Groenigen: Designing Spatial Coverage Samples Using te K-means Clustering Algoritm. Digital Soil Mapping: An Introductory Perspective, Elsevier (006) 4. F.J. Gallego, Stratified Sampling of Satellite Images wit a Systematic Grid of Points, ISPRS Journal of Potogrammetry and Remote Sensing, 005, Vol. 59, pp (005) 5. B. Minasny, A.B. McBratney, and D.J.J. Walvoort: Te Variance Quadtree Algoritm: Use for Spatial Sampling Design. Computers and Geosciences, 007, Vol. 33, pp (007) 6. C. Kadilar, H. Cingi: A New Ratio Estimator in Stratified Random Sampling. Communications in Statistics-Teory and Metods, 005, Vol. 34, No. 3, pp (005) 7. Tetsuji Oyam, J.A.Jimmy A. Doi, and Takasi Yanagawa: Estimating population Caracteristics by Incorporating Prior Values in Stratified Random Sampling/Ranked Set Sampling. Journal of Statistical Planning and Inference, 008, Vol. 38, No., pp (008) 8. C. Kadilar, H.Cingi: Ratio Estimators for te Population Variance in Simple and Stratified Random Sampling. Applied Matematics and Computation, 006, Vol. 73, No., pp (006)

11 9. C.C. Wu, R.C. Zang: Empirical Likeliood Inference Under Stratified Random Sampling in te Presence of Measurement Error. Acta Matematicae Applicatae Sinica, Englis Series, 005, Vol., No. 3, pp (005) 0. S. E. Wrigt, R. B. Noble, and A. J. Bailer: Equal-Precision Allocations and Oter Constraints in Stratified Random Sampling. Journal of Statistical Computation and Simulation, 007, Vol. 77, No., pp (007). V. Bosc, R. Wildner: Optimum Allocation of Stratified Random Samples Designed for Multiple Mean Estimates and Multiple Observed Variables. Communication in Statistics- Teory and Metods, 003, Vol. 3 No. 0, pp (003)