Research On Crop Landscape Model And Its Effects On Crop Identification And Acreage Estimation

Nowadays,regional and country level crop products supply situation seems to be more and more serious due to the the rapid changing of country economic development in China.Crop planting acreage showed to be in an obvious unstable status on the front lines results from the enthusiasm abating of crop planting.Therefore,rapid,accurate,and reliable crop acreage estimation become significant for guidance the macro cropping structure adjustment and control the national agricultural policy analysis.Remote sensing techniques,involved in crop identification and crop acreage estimation to support agricultural decision since 1970 s,now has to assume the responsibility to provide more satisfied crop distribution and acreage estimation results than previous.The key issue to improve the accuracy of crop acreage estimation used remotely sensed images found to be exploring the factors that influencing the accuracy of crop identification and acreage estimation with remote sensing.Many researches found crop classification accuracies alays related to the specific distribution of crops in study area.Therefore,the main contents of the study focus on investigating how to quantitative descript or modelling the crop distribution and how to develop an effective way to improve the accuracy of crop identification and acreage estimation based on the modelIn the study,two individual experimental sits(one in the east of Ontario,Canada,the other in the northeast of China)were selected to be essential to further study the modeling and and effects of crop spatial distribution on crop identification and acreage estimation with remote sensing in three aspects:(1)to analysis the dependence of classification features and classifiers,and preliminary determined the optimal features and classifiers;(2)to analysis the effects of crop landscape on crop identification accuracy,and to propose the concept of crop landscape and a model to quantitative describing crop landscape through four landscape index.Meanwhile,we divides all the landscape units into six types;(3)Optimal classification features,classifiers were determined to the six crop landscape divisions.Crop acreage estimation improving models were also developed and implement on regional rice acreage estimation in Hunan province on the basis of crop landscape division,the results to be tested effective.The main conclusions include the following:(1)Principle component analysis(PCA)binding correlation analysis(CA)and stepwise discriminant analysis(SDA)are feasible and effective for crop identification feature selection with remote sensing.The results showed that: five vegetation indexes,with great contribution on the classification accuracy,are TVI,GNDVI,NDVI,VIgreen and TCARI;seven textural features based on the near infrared band and the 9×9 window size are important for the crop classification.(2)Various combination of the three groups of crop identification features,i.e.reflectance of NVIR bands,22 vegetation index,and GLCM textural features,were evaluated through four classification algorithms(maximum likelihood classifier(MLC),artificial neural network(ANN),support vector machine(SVM)and random forest(RF))and result in different behaviours:(1)the effect on the accuracy of different combination of features is different using different classifiers,there is no uniform optimal classification method;(2)comparing to using band feature only,the including of VIs cannot improve the accuracy well,however,the accuracy increased about 4%-5% when including the textural features.The accuracy reached highest when all the features were used in the classification(increased about 7%);(3)The importance of features using RF-based variable importance method showed that the band feature has the highest contribution on the crop classification,especially the near infrared band,the textures has the lower contribution,and the contribution of VIs is lowest.(3)A conceptual crop landscape model that expressed in four landscape index were proposed to depict the exact crop spatial distribution.Six typical crop landscape divisions were then introduced to stimulate the investigation of relationship between crop spatial distribution and crop identification accuracies.Specifically:(1)the method binding descriptive statistics and normality test,correlation analysis and the method of factor analysis found to be feasible and effective to be used to select landscape indexes to depict crop spatial distribution;(2)the crop landscape model can be expressed by four index,i.e.LSI,FRAC_AM,AI and SHEI;(3)the K-mean clustering method were tested to be satisfied to introducee the landscape units into six typical divisions to support the crop identification and acreage estimation with remote sensing.(4)A lookup table was developed to suggest the optimal classification features and classifiers through experiments at local scale.And the followingwere found that:(1)the VIs feature performed best for crop classification and the band feature performed worst in typical division I;the VIs feature performed best and the textural features has the worst performance in typical division II and IV;the VIs feature performed best and the band and textural features almost the same in typical division III;the band and VIs feature performed almost the same in in typical division V;the band feature performed best and the textural features has the worst performance in typical division VI;(2)The RF performed best in typical division I;the ANN performed best in typical division II,III and IV;the MLC,SVM and RF performed almost the same in typical division V;the MLC performed almost the same in typical division VI;(3)the optimal classification features and classifiers are different with different landscape pattern;the near infrared band in band features have greatest contribution on crop classification accuracy;the GNDVI,TCARI,TVI of VIs and the Mean of textural features are also important on crop classification.(5)At large scale,a series models were proposed to improve crop acreage estimation through local experimental trials.Results showed that there exist good relationship(R2 greater than 0.75)between the acreage estimated two different crop acreage estimation methods(crop estimation through crop landscape divisionand that without crop landscape division),except the division with very low crop proportion and scattering fields.Yet the results with crop landscape division could results in manifest improvement in accuracies.Implementation of the method in Hunan province for rice acreage estimation showed that:(1)the acreage accuracy increased about 6% in the typical division I and II;(2)the accuracy increased 4% in the typical division III,V and VI.The main highlights of the study include the following:(1)A conceptual model that expressed through four crop landscape index was proposed to depict crop spatial distribution.Based on them,the method to support landscape typical division was obtained,and six typical divisions were got according to the method.(2)The effects that crop spatial distribution acts on the optimal crop identification features and classifiers selection were difinited.The best classification accuracy can be obtained at the shortest time under a certain type of landscape.(3)Methods that improve crop identification and acreage estimation were developed based on crop landscape division specifically.The method provided data support and scientific basis on the realization of high precision and fast crop acreage estimation at large scale.