两种水稻种植面积遥感提取方案的分析

A DISCUSSION ON TWO STRATEGIES APPLIED TO ESTIMATE RICE PLANTING AREA OF AN ADMINISTRATIVE DIVISION USING REMOTE SENSING TECHNIQUE

文章以湖北省江陵县的水稻面积提取为例，对这两种方案（分别称之为Ａ方案和Ｂ方案）进行了非监督分类试验。结果表明，在非监督分类过程中Ｂ方案明显优于Ａ方案，所提取的水稻面积精度也在８４％以上。

In crop yield estimation using remotely sensed data, it usually needs to calculate the crop planting area in a particular administrative division. Most previous investigators do as follows: first, they cut down the target image of the study area with the administrative boundary, then conduct land cover/use classification and crop identification work, and finally calculate the crop area. Other researchers conduct the land cover/use classification work first and then cut down the study area with administrative boundary, and calculate the crop area at last. We call these two methods strategy A (cut and classify) and strategy B(classify and cut) respectively. In this paper, we applied these two strategies to rice planting area identification. Our results indicate that strategy B is obviously better than strategy A in the unsupervised-cluster process and the accuracy is over 84%. Previous work The author tried to retrieve as many as possible previous works on land cover/use classification and on crop area calculation in an administrative division. It showed that much of these works was done based on strategy A (cut and classify) using supervised or unsupervised automatic classification method as well as visual interpretation. Strategy B (classify and cut) was used by fewer investigators, in cases supervised classification method was applied. Study area and data The study area we selected is the county of Jiangling in Hubei province, China. Jiangling county, located in the middle Changjiang River Plain, is a major rice production county in Hubei province. The early rice is sowed in the third ten days of March or the first ten days of April and transplanted in the third ten days of April or the first ten days of May. The moderately late rice is sowed in the first ten days of May and transplanted in the third ten days of June. According to the farming practice of the area, Landsat 5 TM CCT, dated 8 June, 1992, when it was clear and cloudless, with scene of Path 124, Row 39 containing the whole county, was acquired from the Chinese Satellite Ground Station. The image processing system we used is ERDAS software and ARC/INFO GIS software is also supplementally used. Moreover, 1∶50000 scale topographic maps, recent vegetation type maps, soil maps, land cover/use maps and other ancillary information were available. Methods and results The process to compare the two strategies was described in detail in this part. For strategy A, similar to previous works, we cut down the image of Jiangling county with its boundary stored in a GIS which was built with ARC/INFO software. Then, unsupervised classification was applied and 50 classes were got at first and then recoded into 10 major land cover types referring to the soil maps and topographic maps. As to strategy B, an image containing Jiangling county was cut down with a circumbox and then go on with unsupervised classification. Fifty classes were got firstly and then recoded into 10 major land cover types also. The classification result was cut down with the same boundary as strategy A. The thematic rice area was extracted using the above two strategies. Results analysis and discussion The results indicated that strategy B is obviously better than strategy A in the unsupervised classification-recoding process. For supervised classification, it made no difference whether strategy A or strategy B was applied. It is expected that some previous visual interpretation results would also be improved if strategy B was used instead of strategy A. The same results can be got for NOAA AVHRR and SPOT images. Conclusions and further research In summary, it may be said that strategy B is far more suitable and robust in the unsupervised classification-recoding process than strategy A. Strategy B, rather than strategy A, should be of first consideration in similar projects. Applying this strategy in our study, the accuracy of the rice area identified is exceeding 84%. In practical use, the rice planting area is always much lager than our study area. A larger crop cultivated area may cover