Abstract:Accurately acquiring crops distribution information is of great significance for agricultural production management and yield estimation, but the roads, forest belts and ditches in the farmland seriously affect the accuracy of crops classification and extraction. Chinese small satellite constellation of small satellites for environment and disaster monitoring and forecasting (HJ-1A/1B satellite) is a good data source for crops classification, because it is free for researchers and has a higher spatial resolution of 30m and a higher time resolution of two days. In this paper, Shuanghe farm in Heilongjiang province of China was the research area, 23 timeseries HJ-1A/1B images which cover the growth period of the major crops from April 3th to November 9th, 2012, were used to monitor the roads and forest belts in the farm, extract spatial distribution of the major crops based on decision tree and objectoriented method, and the classification result was compared to traditional decision tree. The timeseries image set and the timeseries characteristic index set such as NDVI, DVI, RVI, EVI and NDWI were built after the original image data pretreatment. Firstly, the road in the farm was extracted with objectoriented classification based on elements of lengthwidth ratio and other parameters, then the timeseries set was masked by the road in order to rule out the interference of roads, forest belts and ditches for the extraction of crops information. Secondly, seven effective characteristic parameters and 14 sensitive time phases were chosen by using the object spectrum, time phase and time series characteristics. The thresholds of characteristic parameters were determined, and the decision tree classification model of major crops was established. Finally, the major crops in Shuanghe farm such as corn and rice were extracted. The result showed that using many characteristic indices to classify crops was very effective, and especially NDWI was very helpful for rice extraction. The method of decision tree and objectoriented classification was better than the traditional decision tree for extracting the spatial distribution of major crops in Shuanghe farm, it could effectively eliminate the interference of roads, forest belts and ditches in the farm for crops classification, and the total accuracy was increased from 89.22% to 95.18%. The integration of decision tree and objectoriented classification can provide reference for crops distribution information extraction in other agricultural areas with low cost and high precision.
