关键词: 时间序列遥感影像, 作物分类, 时效性, 决策树, 最大似然法

Abstract: 【Objective】Crop type remote sensing identification is a basis of crop cultivated area and crop growth analysis and yield estimation, and it is a very important driving force to promote the rapid development of modern agriculture. At the same time, it is also a basis for macro-regulation and control of understanding of agricultural conditions by the departments of agriculture as well as other related ones. At present, most of the present researches about agricultural remote sensing are limited to moderate or low resolution remote sensing images, which affect the accuracy of vegetable information extraction. The accuracy of vegetation information extraction can be improved by using high resolution multi temporal remote sensing images and selecting suitable classification methods. Clearly understanding of the timeliness and optimal classification method of crop remote sensing classification, acquire crop spatial distribution data quickly and accurately, and to provide a basis for crop quantitative remote sensing monitoring are the aims of the study.【Method】Based on the 20 remote sensing images covering the whole growth period of 5-10 months in Hulin, Heilongjiang province in 2014, the 16 m resolution NDVI time series curves were built by using 20 images. Different crops had different NDVI time series curves during the whole growth period. The decision tree classification model was established. After analysis of the images through serial threshold division, assisted with background data and expert knowledge, the areas and distributions of the land use and land cover information were extracted. Twenty images were used in order to classify the crops and the optimal phase was defined. Taking the farmland range as the rule, various classification methods for crop classification were compared. And it was also compared with the crop classification without extracting the farmland range by using several common methods of crop classification. Meanwhile, various classification methods including the maximum likelihood method, Mahalanobis distance method, neural network method, minimum distance method, support vector machine, spectral angle classification, and crop classification of principal component analysis were compared, and the data from the insured blocks were employed for the accuracy verification.【Result】 (1) In early July, the end of July to early August, and the end of September are the 3 key phases of crop remote sensing classification in the study area during the first quarter of the year. (2) The decision tree classification method had the highest accuracy in extracting land use cover information, the overall accuracy of classification was up to 94.01%, Kappa coefficient was 0.79. (3) In early June and early July, 2 images combined with classification of crops, the overall of classification accuracy was up to 90.24%, Kappa coefficient was 0.87. The combination of early June and early July images could be used to solve the timeliness of crop classification. (4) Combined with the image of Sep 21st, the overall accuracy was further improved, and the classification accuracy of soybean was improved obviously, so the maximum likelihood method was the best classification method, and the jointing stage was the best phase.【Conclusion】It was concluded that remote sensing images can be used to accurately classify crops in early July. Results of this study have expanded the application value of remote sensing data in the field of agriculture. It has guiding significance for one crop a year of the crop fast classification.

Key words: time series remote sensing image, crop classification, timeliness, decision tree, maximum likelihood method