基于光谱时间序列拟合的中国南方水稻遥感识别方法研究

多云多雾现象是农作物遥感分类经常遇到的问题,影响分类精度。为解决此类问题,本文提出一种基于时间序列GF-1号遥感影像识别水稻方法。利用多时相时间序列的GF-1号遥感影像提取中稻、晚稻的近红外波段（NIR）反射率、红光（R）波段反射率、归一化植被指数（NDVI）特征;拟合光谱和植被指数时间序列特征曲线;分析多时相影像离散近红外波段 、红光波段、NDVI值落在拟合中稻、晚稻近红外波段、红光波段、NDVI时间序列曲线两侧的敏感性区域的比例,该区域也可以视为水稻作物识别特征的目标特征区域,只有达到一定的比例才能视为某类水稻作物。在此情形下,需要综合3种情况进行集中投票决定其最终分类结果。研究表明：该方法可以在多云雾地区对中稻和晚稻精确识别,中稻和晚稻用户精度可达95.97%和95.95%,总体精度为95.76%,kappa系数为0.9335。实验结果表明了NIR、R、NDVI时间序列曲线拟合的有效性,以及拟合曲线目标特征区域设置的合理性。

Research on the Method of Rice remote sensing Identification Based on SpectralTime-series Fitting in Southern China

Food security is an important guarantee for the stable development of our country and the area of planting grain is the basis of food security, so the estimation of the area of planting grain is important. Remote sensing technology is an important method of estimating crop grain area at present. The classification accuracy is affected by cloud and mist, which cannot be avoided. To solve this problem, this study presented a method for recognizing rice based on GF-1 time-series image. With long time-series of GF-1 images, three indices of middle-season rice and late-season rice, namely near infrared band reflectance (NIR), red (R) band reflectance and the normalized difference vegetation index (NDVI) characteristics are extracted. Spectrum and the characteristic curve of vegetation index time-series are fitted. We analyzed the ratios of values of discrete near infrared band, red light band and NDVI of images of multiple temporal phases falling on both sides of the sensitive area of the fitting NIR, R and NDVI time-series curve of middle-season rice and late-season rice. This area can also be seen as the target area of rice identification features and only those reaching a certain proportion can be identified as certain type of rice. Under this condition, three kinds of situation should be considered comprehensively and voted to decide final classification results. The means of samples are used to fit the curve for each image. The outliers are eliminated from the ground samples in advance. Statistical analysis of ground samples defined target characteristics. The result indicated that: (1) Using polynomial fitting method based on least square principle to fit NIR, R, NDVI time series characteristic curve, fitting effect is better when fitting degree is 3 and it can satisfy the need of subsequent classification. (2) Different setting proportions led to different classification accuracy, and the overall accuracy is 95.76%, the user accuracy of middle-season rice and late-season is 95.97% and 95.95% when the setting proportion is not less than 50%. (3) The method proposed in this study could solve the problem of the combination of complex phases, and significantly weaken the influence of cloud and fog on crop classification, especially in South China.