2022, 33(2): 218-230.
DOI: 10.11898/1001-7313.20220208
Abstract:
Various high-resolution satellite remote sensing platforms and sensors have been established recently, but there are still many challenges for crop type fine classification. New methods are needed for fine classification of crop types and extraction of peanut planting area using high-resolution optical remote sensing data. Using panchromatic and multi-spectral (PMS) data of GF-1 and GF-6 in the middle and late stages of crop growth, 40 classification features are constructed from the perspective of spectrum, texture, and temporal variation of remote sensing image by means of spatio-temporal fusion and spectral mathematical transformation.15 features are selected by ReliefF-Pearson method. To compare the effects of different feature types on classification, the selected features are combined into four feature spaces, and the separability between crops in these feature spaces is measured by Jeffries-Matusita (J-M) distance. Using maximum likelihood classification, support vector machine and random forest as classifiers, crop classification tests of four feature space schemes are carried out, and different classification results are evaluated from the perspective of remote sensing classification accuracy and landscape analysis. Compared with the original image, in four feature spaces, the J-M distance between peanut and other crops is close to the saturation value of 2.0, and their separability is significantly enhanced. For all test results, the overall accuracy of crop classification is more than 78.0%, the Kappa coefficient of crop classification is more than 0.7, and the user's accuracy and producer's accuracy of peanut are both more than 79.0%. Compared with the three classifiers, Random Forest has the highest classification accuracy, and the average values of the overall accuracy and Kappa coefficient, peanut user's accuracy and producer's accuracy of its four schemes are 91.49%, 0.89, 94.45% and 93.30%, respectively. In addition, combining dual-temporal remote sensing data, texture and vegetation index features, as well as their temporal changes, can effectively improve the classification accuracy. It shows that adding the texture feature can improve the accuracy better than the vegetation index features. The landscape shape analysis of the classification results also shows that the overall crop classification and peanut, their patch shapes are close to the reality, and random forest is the best classifier. Its mean shape indices of crop landscape-level and peanut class-level are only 1.333 and 1.270, and the corresponding mean patch fractal dimension indices are only 1.127 and 1.110. An optimal dual temporal remote sensing crop classification model is proposed. Using this model, summer peanut planting area extraction and area measurement are carried out in four main peanut producing counties in Huang-Huai-Hai Region. Compared with the statistical data, the determination coefficient is 0.98 and the relative error of area calculation is ±16.25%, showing a good application prospect.
Guo Qile, Li Junling, Guo Peng. Extraction of peanut planting area based on dual-temporal remote sensing features of crops. J Appl Meteor Sci, 2022, 33(2): 218-230. DOI: 10.11898/1001-7313.20220208
