中低分冬小麦分布提取模型效率的样本特征分析

中空间分辨率样本（简称中分样本）的数量、质量是决定中低分辨率复合识别模型效率的关键因素。以冬小麦为研究对象，中低分辨率影像结合构建支撑向量回归模型（Support Vector Regression,SVR）实现冬小麦的混合像元分解，提取出冬小麦的空间分布，定量分析中分样本数量、质量对识别精度的影响。结果表明：从样本数量上看，样本量为10%即可保证稳定的冬小麦精度，在典型冬小麦区的区域精度、像元精度达到98%、92%以上；从样本质量上看，识别精度随样本质量（达到60%即可获得较好的识别结果）增加而升高；对于非中分样本区的冬小麦，区域精度、像元精度也是随样本数量的增加而提高，在20%样本量下，区域精度和像元精度稳定在97%、92%以上，表明该模型具有较强的空间泛化性能力，弥补了从低分影像上难以获取有效样本的不足。

Influence Factors Analysis on Accuracies of Winter Wheat Distribution from Low and Medium Resolution Composited Remote Sensing Images

The quality and quantity of sample dataset from medium resolution remote sensing images is the key factor to contribute to the efficiency of low and medium resolution identification model. For winter wheat in this paper， we constructed a support vector regression model coupled with low and medium resolution images， to decomposed of mixed pixels， and exact winter wheat extent. Then analyzed the influences of sample quantity and quality of medium resolution remote sensing images respectively. The results states that only 10% quantity of samples are enough to achieve stable accuracy. Under this quantity， regional accuracy and pixel accuracy could reach higher than 98% and 92% respectively in typical winter wheat area. In terms of sample quality， the accuracy of result improved accompanying with the sample quality increment. We found that high accuracy could achieved when the sample quality is better than 60%. While in the area where medium resolution sample did not exist in area with medium samples， regional accuracy and pixel accuracy also increased accompanying with the sample amount and quality increment. In this area， 20% quantity of medium resolution sample was needed enough to achieve 97% of regional accuracy and 92% of pixel accuracy respectively. The above demonstrate the successful generalization of winter wheat identification by medium resolution sample to non-medium resolution area.