1.Classification of crops by remote sensing
1.1 Crop classification and identification experiment using multi-temporal environmental satellite CCD data
The support vector machine classifier was used to classify images based on pixels. In the classification process, the single-scene environment star data on different days and the combination of environment star data on different days were classified respectively to evaluate the environment star presencecropsThe application potential of classification, and determine the optimal image acquisition period and optimal time combination for crop classification using environmental star data.
The results of classifying the environment star data for a single phase and a combination of different phases are shown in Figure 1.
The lower left corner ** indicates the environment star data used for classification, for example, hj3 indicates that the environment star data obtained in March is used, and hj3+hj4 indicates that the combination of environment star data obtained in March and April is used.
The classification effect of environmental star data obtained during the flowering period was the best, and the overall classification accuracy reached 882%, the data obtained in the early stage of flowering was the second, and the data obtained at the jointing stage had the worst classification effect among the images of the three phases. The classification effect of data using multiple phases is better than that of data using single phases, and the data classification effect of using three phases is the best, and the overall classification accuracy reaches 917%。However, the use of data from three phases has limited improvement in classification accuracy compared with the combination of two phases at flowering and jointing stages, indicating that the combination of two suitable phases can achieve sufficient classification accuracy, and the combination of more phases has little effect on the improvement of classification accuracy.
1.2 Crop classification and identification experiment based on the fusion of environmental star and ASAR data
Data fusion: Principal component analysis (PCA) was used to fuse the environment star and ASAR data. The environmental star data containing four different spectral bands were transformed by principal component transformation, and the ASAR VV polarization image was stretched to make the mean value of the image gray scale consistent with the first component image of variance and principal component transformation, and then the first component image after the principal component transformation of the environment star data was replaced with the stretched ASAR image, and then restored to the original image space through inverse transformation to obtain the fusion data of the environment star data and the ASAR data. The spectral angle was used to measure the spectral difference between different ground objects before and after fusion.
The classification algorithms used in this study include Maximum Likelihood (MLC), Artificial Neural Network (ANN) and Support Vector Machine (SVM). Compared with the environment star data, the image level after the fusion with the ASAR VV polarization data is more distinct, and the differences between different ground objects are more obvious. Although the spatial resolution of the ASAR data is the same as that of the environment star data, the boundaries of the fused fields are more obvious than those of the environment star because the ASAR VV polarization data is more sensitive to the field boundary.
Fig. 2 Comparison of the visual effects of the field boundary on different image products Left: Environment star;Medium: Blend image;Right: rapideye
The spectral angles between the categories increased significantly after fusion, indicating that the spectral differences between different classes increased significantly after fusion. The distribution of spectral angular values between wheat and cotton, wheat and trees, and cotton and trees has been changed from 0063、 0.02、 0.043 Enlarged to fused. 104。Compared with the ASAR VV polarimetric backscatter data, the amount of information contained in the environment star multispectral data is significantly increased, which expands the spectral difference between different ground objects and enhances the separability between ground objects.
According to the ground survey and RapidEye visual interpretation, the random sampling method was used to select the classification samples. Half of the ground samples were randomly selected as the training samples, and the other half were used as the test samples for accuracy evaluation, and the training samples and the test samples did not coincide. Different classifiers were used to classify the environment star data and the fused environment star and ASAR data, respectively, and the results are shown in Figure 3.
A, B, and C are the classification results of the environment star data using the MLC, ANN and SVM methods, respectively, and D, E, and F are the classification results of the fusion data using the MLC, ANN and SVM methods, respectively.
From the visual effect, the environmental satellite multispectral data can effectively identify the wheat planting plots, and the support vector machine classification effect is the best, but some cotton plots are missing. After the fusion of environmental star multispectral data and ASAR VV polarization backscatter data, wheat plots can be effectively identified, and the field boundary is more obvious, and the classification effect is better than that of environmental star multispectral data alone.
In terms of classification methods, the support vector machine classification method is the best for both ambient star data and fusion data, and the classification accuracy of using fusion data is about 5 percentage points higher than that of environment star data alone, reaching 943％。
In summary, the environmental satellite multispectral data can be used to classify crops effectively, but there are problems that the field boundaries cannot be effectively identified and the classification is confusedThe ASAR VV polarization data can improve the spectral information of the optical data, so that the spectral difference between different ground objects can be significantly increased, and the separability between ground objects can be enhancedVV polarization data is sensitive to field non-cultivated land information, and has a great effect on the identification of field boundariesThe VV polarization data is sensitive to the structure information of ground objects, which leads to a slight expansion of the field boundary, resulting in a slight decrease in the proportion of wheat area in the classification results, but it is worth improving the classification accuracy.
a.Crop classification and identification experiments using multi-source SAR data.
During the 2009 winter wheat regreening period, ASAR VV polarization data of three growth stages of wheat regreening, jointing and flowering were obtained, with a spatial resolution of 30 m and C-band data. The data received on -4-3 and 2009-5-8. At the same time, the data of Terrasar-X was obtained, with HH polarization, spatial resolution of 6 meters, and the data received at the flowering stage of wheat on May 10, 2009. The remote sensing image classification method using support vector machine classifier is used. In the process of classification, different combinations of SAR data were classified to evaluate the application potential of multi-source SAR data in crop classification and determine the optimal combination of SAR data for crop classification. The results of the four-scene data combination are shown in Figure 4.
Figure 4 Classification results for a combination of 3-scene ASAR data + 1-scenario Terrasar data.
The accuracy evaluation results show that the overall classification accuracy is low when using data of a single phase for classification. The classification effect of data using multiple phases is better than that of data classification of single phase, and the classification effect of data using three phases is the best, and the overall classification accuracy reaches 8412%。The classification accuracy of SAR data at two frequencies is better than that of multi-phase data, and the classification accuracy of A3+T reaches 8655%, which is higher than the ASAR classification accuracy for the three phases. After adding texture information, the classification accuracy has been improved to a certain extent compared with the backscatter data, about 3-5 percentage points.
Through the classification experiments of multi-source SAR data, the following conclusions are drawn: (1) The combination of multi-frequency SAR data obtains higher classification accuracy than multi-temporal data. (2) The combination of the two temporal ASAR data at jointing stage and flowering stage obtained the same classification accuracy as that of the three temporal ASAR data. (3) The addition of texture information can improve the accuracy of SAR crop classification.
2.Crop acreage estimation from multi-scale remote sensing data
The experimental area is selected for the northern part of the North China Plain, which spans the central and southern parts of Hebei, the northern part of Shandong, and the northern part of Henan, covering an area of about 200,000 square kilometers.
The medium-resolution remote sensing data mainly includes 14-period modis NDVI 16-day synthetic dataThe high-resolution remote sensing data is based on CBERS CCD data, mainly 6 scenes.
The technical methods used are: 1) the use of high-resolution imagery for crop classification with the support of ground survey data and **g data;2) The classification results were counted in the region, and the area size was used in the same large grid as the medium and low-resolution image pixels, and the results of crop components with the same pixel size as Meris or Modis were obtained3) Neural network model was established by comparing the Modis time series NDVI dataset with Envisat Meris multispectral data with high-resolution generated crop component data4) The neural network model was used to extrapolate the whole image to obtain the crop planting area in the study area.
The experimental results show that the accuracy of estimating maize planting area by using multi-scale remote sensing image data can reach 90%, especially at the level of the whole study areaAt the prefectural and municipal levels, the accuracy of maize planting area estimation using MDOIS NDVI can reach more than 90%, and at the provincial scale, the accuracy of estimating maize planting area using MDIS NDVI can reach more than 95% in Hebei and Shandong, and the estimation accuracy of MERA can only reach more than 90%.