利用物候差异与面向对象决策树提取油菜种植面积

及时、准确地获取农作物种植信息,对于农业生产管理和国家粮食安全有重要意义。目前越来越多的免费卫星数据可以用于作物分类及生理参数反演。Sentinel-2卫星于2015年6月发射,提供了13个光谱波段,具有较高的时间分辨率、空间分辨率和光谱分辨率,为不同作物特征区分以及大范围作物种植面积快速提取业务化运行的精度与效率提高带来了契机。随着Sentinel-2数据的免费下载,这就为大面积生产下一代区域或者国家尺度的高分辨率（10~30 m）农情遥感产品提供了可能。物候信息包含了作物随着季节不断变化的特征,利用如NDVI等时间序列植被指数找出不同作物的特征进而开展作物分类得到了广泛应用。本文以油菜为主要研究对象,以长江中下游地区的江汉平原为实验区,基于作物物候差异与面向对象决策树的方法,对Sentinel-2卫星影像用于油菜种植区提取的效果进行了评估与分析。首先利用作物不同生长时期各波段光谱信息以及归一化植被指数等信息的差异分析并找出油菜种植区提取的最佳时相,然后对影像进行多尺度分割,根据对象特征建立决策树逐一去除非植被、林地等干扰类型,进而提取出油菜种植区域。通过分析发现,基于Sentinel-2影像的图像分割可以有效生成不同作物类型的对象;油菜开花期的特征是其区分于其他作物的关键因素,利用该特征可以有效消除分类时其他地物类型对油菜的影响,提高作物分类信息提取的精度和效率。研究表明：在区分油菜的决策树分类特征信息中,贡献最大的是归一化植被指数（NDVI）,近红外波段（NIR）和亮度（Brightness）信息。用162个油菜验证样本点计算混淆矩阵,油菜种植面积提取的总体分类精度为98%以上,Kappa系数为0.95。说明结合物候信息利用Sentinel-2数据进行大范围作物种植面积提取具有巨大潜力,可以提高大范围油菜种植区域快速提取的精度和效率。

Identifying Rapeseed Planting Area Using an Object-oriented Method and Crop Phenology

Accurate and timely access to crop planting information is important for agricultural production management and national food security. Increased access to free satellite data, has allowed for further crop classification and physiological parameter inversion. Launched in June 2015, the Sentinel-2 satellite provides 13 spectral bands with high temporal, spatial and spectral resolution. This provides an opportunity to improve the operational accuracy and efficiency of rapid extraction of crop characteristics and large-scale crop planting area. The Sentinel-2 satellite provides free images at a high spatial resolution (10~30 m). Therefore, it is now possible to develop the next generation of agricultural product at both national and regional levels. Time series vegetation indices (VIs) are derived from this satellite data and are widely used to identify crop characteristics for classification. The Jianghan Plain (located in the middle and lower reaches of Yangtze River) was used as a case study area to evaluate and analyze the effect of the Sentinel-2 satellite images on rape planting area, based on an object-oriented method and differences in crop phenology. Firstly, the best estimate of rapeseed planting area was identified by using crop spectral information and normalized ve getation index at different growth stages. Secondly, a decision tree was formulated according to the difference in object features after multi-scale segmentation. Finally, the rapeseed planting area was calculated after the removal of non-vegetation areas, forests, winter wheat crops and other interferences based on the decision tree. It was found that image segmentation based on the Sentinel-2 images could identify various crop types effectively. The flowering characteristic of rapeseed is a key factor in distinguishing it from other crops. The difference in object features can effectively eliminate the effects of other land-use types on rapeseed classification and improve the accuracy and efficiency of crop classification. Results showed that the greatest classification contributors to the decision tree for distinguishing rapeseed were the normalized vegetation index (NDVI), near infrared (NIR), and brightness (Brightness). A total of 162 rapeseed sample points were used to calculate the confusion matrix. The overall classification accuracy of the rapeseed planting area was more than 98%, and the Kappa coefficient was 0.95. This shows that the Sentinel-2 data, combined with phenology information has great potential to extract large-scale crop planting area data, and can improve the accuracy and efficiency of calculating rape planting areas at both national and regional scales.