Extraction Methods Of Orchards And Facility Agriculture Land Based On High Spatial Resolution Remote Sensing Images

Precision agriculture has been the main development direction of modern agriculture in China.Extracting the spatial distribution information of orchards and facility agriculture land from high spatial resolution(HSR)images rapidly and accurately is of great significance for the precise management,the optimization of industrial layout and the sustainable development of the orchard and facility agriculture.In view of the problems to be solved urgently in identifying the agricultural land with remote sensing images in the present study,such as fewer researches on the extraction of orchards and facility agriculture land compared with other croplands,the spectral confusion problems in HSR images and the low classification accuracy of the current methods,with the main objective of improving the extraction accuracy of orchards and facility agriculture land in small scale area,based on the QuickBird,SPOT-7 and GeoEye-1 satellite image data,the key techniques and principles such as the spectral distinguishability of different ground objects in HSR images,the extracting methods of different texture features,the selection of optimal classification features,the construction of supervised classification models,the preprocessing of the HSR images and so on were studied in this paper.The main conclusions were as follows:(1)In order to clarify the spectral characteristics of the HSR images used and the spectral distinctiveness of the typical ground objects,the classifications based on different spectral features and the SVM algorithm were carried out.The spectral statistical results of different ground objects showed that there were some spectral confusions between vegetation types themselves and non-vegetation types themselves in the study area.The statistical results of the four multi-spectral bands showed that the near-infrared band had the largest mean and standard deviation,and its correlations with the other three spectral bands were the least,which indicated that the brightness of the near-infrared band was the largest,the corresponding variations were obvious,the information content within the near-infrared band was the most abundant and also the information overlaps with other bands were the least,all of which were helpful to increase the spectral discriminability among different ground objects.The conventional classification results based on the SVM algorithm showed that the accuracy based on spectral-only features with different HSR images were generally low,and the contribution of the near-infrared band to the classification was higher than that of the other three bands.The highest OA and F_a(the extraction accuracy of the apple orchard)in the apple orchard study area were 86.62%and 85.05%,respectively.The highest OA and F_k(the extraction accuracy of the kiwifruit orchard)in the kiwifruit orchard study area were 83.37%and 79.90%,respectively.The highest OA and F_g(the extraction accuracy of the greenhouse)in the facility agriculture study area were 85.91%and 82.32%(SPOT-7 image),92.22%and91.27%(GeoEye-1 image),respectively.(2)The orchards showed remarkable texture characteristics in the HSR images.Aiming at the spectral confusion between apple orchards and other vegetation types,kiwifruit orchards and other vegetation types in the specific study area,the texture feature extraction methods based on the GLCM method,the wavelet transform,the fractal model and the spatial autocorrelation analysis were studied respectively.According to the planting characteristics and image features of the apple orchard land,a texture extraction method based on the GLCM analysis was designed.The results showed that the GLCM texture could effectively improve the F_a and OA,and the SF+GLCM TF features achieved the highest F_a and OA in the apple orchard study area,which was 14.17%and 12.39%higher than that of the spectral only(SF)classification results in terms of F_a and OA,and slightly higher(0.63%and 1.56%respectively)than that of the SF+fractal TF classification,but significantly higher(11.92%and 9.20%respectively)than that of the SF+correlation TF classification.The F_a and OA of the SF+GLCM TF and the SF+fractal TF were both higher than 94%,which indicated that both texture features could identify the apple orchards well.The cultivation characteristics and image features of the kiwifruit orchards were analyzed,and a texture extraction method based on the two-dimensional discrete wavelet transform was proposed.The coif5 wavelet function was used to decompose the QuickBird panchromatic image with a two-level wavelet transform,and the energy characteristics of the wavelet coefficients was computed as the texture feature.The results showed that the F_k could reach 87.61%when the wavelet texture(wavelet TF)was used alone,which was obviously superior to the spectral only classification and the other two kinds of texture only classifications.The F_k and OA of the SF+wavelet TF classification were the highest and both higher than 94%,which was 15.03%and 8.94%higher than that of the spectral only classification in terms of F_k and OA,6.70%and 2.88%higher respectively than that of the SF+GLCM TF features,also 13.43%and 6.98%higher respectively than that of SF+fractal TF features.The results showed that the wavelet textures could effectively increase the distinguishing ability between different ground types,the SF+wavelet TF features achieved the best results in mapping kiwifruit orchards among all the features used.(3)In order to improve the extracting accuracy of the apple orchard land in the QuickBird images,an automatic and comprehensive method based on spectral features,GLCM texture features and the SVM algorithm(SF+GLCM TF_SVM)was proposed.This method firstly computed the GLCM texture features from the QuickBird panchromatic image,then constructed the classification features by combing the GLCM textures and the spectral features,and finally delineated the apple orchards using the combined features and the SVM classification.The experimental results showed that the F_a and OA of the SVM classification were higher than that of the MLC results and the ANN results under the same classification features(SF,SF+GLCM TF,SF+fractal TF,SF+correlation TF).The SF+GLCM TF_SVM model achieved the highest classification accuracy with the F_a and OA of 96.99%and 96.16%respectively.Compared with the SF+GLCM TF_MLC model,the F_a and OA increased by0.93%and 2.57%respectively.Compared with the SF+GLCM TF_ANN model,the F_a and OA increased significantly by 14.67%and 8.78%respectively.This showed that the apple orchard detecting model based on the SVM algorithm was better than that of the MLC and the ANN.Consistency of the extracted apple orchard area and the visual interpretation results were able to achieve 98%in the test regions.(4)Aiming at the problems in the effective identification of the kiwifruit orchard land in complex planting environment of multiple orchards co-planting,a hybrid method(SF+wavelet TF_RF)for the detection of the kiwifruit orchard land based on the wavelet transform and the random forest algorithm was proposed.The method firstly computed the wavelet texture features from the QuickBird panchromatic images,then constructed the classification features by combining the wavelet textures with spectral features,finally the kiwifruit orchard distributions were automatically delineated by the random forest(RF)ensemble classification technique.The results showed that the RF results were superior to those of the SVM and the MLC results with the same classification features(SF,SF+GLCM TF,SF+fractal TF,SF+wavelet TF).The SF+wavelet TF_RF extracting model achieved the highest classification accuracy with the F_k and OA of 95.30%and 94.46%respectively.Compared with the SF+wavelet TF_SVM classification,the F_k and OA increased by 1.65%and 3.38%respectively.Compared with the SF+wavelet TF_MLC classification,the F_k and OA increased by 6.07%and 4.13%respectively.An apple orchard extracting experiment was also carried out using the same method,and the results indicated that the method had good applicability,with the F_a and OA of 97.77%and 96.43%respectively.(5)An effective mapping model for the facility agricultural land integrating the spectral information,the wavelet texture information of the HSR images and the RF algorithm was constructed.The model was used to delineate the greenhouse distribution from the GeoEye-1and SPOT-7 images.The results showed that the OA and F_g of the images were both higher than 85%with the spectral only features.When the spectral features and the texture features were combined,the maximum F_g and OA of the GeoEye-1 image and the SPOT-7 image reached 94.29%and 94.58%,92.67%and 92.52%respectively,which indicated that both images could be used to identify the greenhouses effectively,and the GeoEye-1 images were superior to the SPOT-7 images in terms of the extraction accuracy.As to the classification features,the combination of spectral features and the wavelet textures could effectively improve the greenhouse recognition accuracy.The SF+wavelet TF classification of the SPOT-7 images was increased by 8.21%and 5.47%,2.11%and 1.82%respectively than that of the SF classification and the SF+GLCM TF classification in terms of F_g and OA.The F_g and OA with the SF+wavelet TF features of the GeoEye-1 images was 2.01%and 1.97%,1.16%and1.06%higher respectively than that of the spectral only classification,the SF+GLCM TF classification.In the selection of classification methods,the RF algorithm had better classification accuracy under the three different kinds of classification features of both images,and the enhancement range of the F_g was between 1.01%and 4.74%,which indicated that the RF algorithm had better classification accuracy and stability than the SVM algorithm.