统计遥感技术是我国统计部门进行国家级、省级和县级作物面积调查的技术手段，该方法集成卫星遥感全覆盖和空间抽样技术，结合地面样方进行区域面积的校正。准确的抽样村作物面积提取是推断多级区域作物面积的关键。目前，抽样村作物面积仍采用人工数字化的手段进行提取，人为主观因素影响大、标准执行难以统一、工作效率低，因此亟需探讨一种自动化手段进行村级作物分布的提取，提高统计调查的效率。

本文在湖北省选了平原地区、丘陵地区和山区三种不同景观的抽样村，利用支持向量机分类器进行了单期、两期影像中稻种植分布提取，进而以最优的像元中稻分类提取结果为基础，分析在不同农业景观下纯净地块最优划分阈值，实现地块级别的中稻自动化提取，最后对比分析不同分类方法对中稻提取的适用性。主要研究结论：

（1）基于像元分类提取中稻面积，根据单期、两期影像的不同组合，可以保证中稻的识别精度，消除同期作物的干扰。在种植结构简单的大坪村采用单期影像可获得较高分类精度，中稻分类用户精度达到100%。对于种植结构复杂地区，两期影像进行中稻的提取，可以有效地消除同期作物的光谱特征干扰，万水桥村中稻分类用户精度从单期影像92.9%提高到95.9%，袁码头村中稻分类用户精度从单期影像93.3%提高到95.0%。但三个村中稻面积精度为92%-93%，还无法达到统计调查精度的要求，这主要是像元分类易受到地物像元光谱异质性、混合像元的干扰。

（2）基于地块中稻提取方法提高了村级中稻的识别精度，一定程度消除光谱和混合像元的干扰。基于地块的中稻分类，以地块为对象进行分析，迭代分析地块内识别中稻像元的丰度确定阈值，划分纯净、混合中稻地块，得到最终的中稻识别结果。万水桥村、袁码头村和大坪村的中稻面积精度分别达到96.1%、97.2%和99.8%，较像元最优分类精度提高了3.0%、3.7%和7.5%。研究发现，纯净中稻地块提取阈值与区域景观特征有着直接的关系，根据地块规整与破碎程度排序，依次为万水桥村75%、大坪村60%、袁码头村50%，这些阈值在一定程度上可以迁移应用到相似农业景观的区域。

       通过在三种不同农业景观进行中稻识别，基于地块的中稻识别面积精度稳定在95%以上，达到了统计调查的业务需求，识别效率较人工数字化方法有着明显的优势，突破了农业统计遥感中抽样村调查的核心技术，提高了整个业务调查流程的效率。该方法为在全省、甚至全国开展抽样调查村的农作物精细化识别打下了实验基础。

Statistical remote sensing technology is a technical means for the national, provincial and county-level crop area surveys conducted by the statistical department of our country. This method integrates satellite remote sensing full coverage and spatial sampling techniques, combined with the ground sample to correct the area. Accurate extraction of crop area in sampled villages is the key to inferring crop area in multi-level areas. At present, the crop area of the sampled villages is still extracted by manual digital methods. The influence of human subjective factors is large, the implementation of standards is difficult to unify, and the work efficiency is low. Therefore, it is urgent to explore an automated method to extract the distribution of crops at the village level to improve the statistical survey effectiveness.

In this paper, three sample villages with different landscapes in plain, hilly and mountainous areas were selected in Hubei Province. Using support vector machine classifiers, single-phase and two-phase images of mid-season rice planting distribution were extracted. Based on the results, the optimal division thresholds of pure parcels in different agricultural landscapes are analyzed to realize the automatic extraction of mid-season rice at the parcel level. Finally, the applicability of different classification methods for mid-season rice extraction is compared and analyzed. Main research conclusions:

(1) Based on pixel classification, the area of mid-season rice is extracted. According to different combinations of single-phase and two-phase images, the recognition accuracy of mid-season rice can be guaranteed and the interference of crops in the same period can be eliminated. In Daping Village with a simple planting structure, single-phase image can be used to obtain a higher classification accuracy, and the accuracy of the mid-season rice classification user reaches 100%. For areas with complex planting structures, the extraction of mid-season rice from the two-phase image can effectively eliminate the interference of spectral characteristics of crops during the same period. The accuracy of mid-season rice classification in Wanshuiqiao Village has increased from 92.9%(based on single-phase image) to 95.9%. The accuracy of mid-season rice classification in Yuanmatou Village has increased from 93.3%(based on single-phase image) to 95.0%. However, the accuracy of the mid-season rice area in the three villages is 92%-93%, and the accuracy of the statistical investigation cannot be met. This is mainly because the pixel classification is easily interfered by the spectral heterogeneity of the feature pixels and mixed pixels.

(2) The method of extracting mid-season rice based on parcels improves the identification accuracy of village-level mid-season rice, and eliminates the interference of spectrum and mixed pixels to some extent. Mid-season rice classification based on parcel, the parcel is used as the object for analysis. Iteratively analyzes the abundance of the identified mid-season rice pixels in the parcel to determine the threshold, divide the pure and mixed mid-season rice parcel and obtain the final mid-season rice recognition results. The accuracy of mid-season rice area in Wanshuiqiao Village, Yuanmatou Village and Daping Village reached 96.1%, 97.2% and 99.8% respectively, which was improved by 3.0%, 3.7% and 7.5% compared with the optimal classification accuracy of pixels. The study found that the extraction threshold of pure mid-season rice parcels has a direct relationship with the regional landscape features. According to the order of the parcels and the degree of fragmentation, it is in order of 75% in Wanshuiqiao Village, 60% in Daping Village, and 50% in Yuanmatou Village. To a certain extent, it can be migrated and applied to areas with similar agricultural landscapes.

        Through the identification of mid-season rice in three different agricultural landscapes, the accuracy of the mid-season rice identification area based on parcels is stable at more than 95%, which meets the business needs of statistical surveys. The identification efficiency has obvious advantages over manual digital methods, which breaks through The core technology of sampling village investigation has improved the efficiency of the entire business investigation process. This method has laid an experimental foundation for the refined identification of crops in sample survey villages in the whole province and even the whole country.