基于低空无人机遥感技术的油菜机械直播苗期株数识别

摘要

Identification of plant number of rape seedlings is the key to evaluating effect and quality of mechanical planting. In terms of the long-lasting and inefficient manual statistics in plant quantity detection of rape seedlings planted mechanically, this article explored a new means to automatically identify the number by a low altitude unmanned aerial vehicles (UAV) remote sensing system with ultra-high resolution. A color vegetation index, excess green - excess red (ExG-ExR), was chosen for image segmentation which was performed by Otsu algorithm. The features of rape plant like spectral and shape information were extracted after image segmentation. Combined with the field survey data, the research applied stepwise multiple regression analysis to build the relationship between the plant number of rape seedlings and features. The low altitude UAV used in research was integrated with an ultra-high resolution sensor, Nikon D800, an FX-format digital single lens reflex (DSLR) camera with effective pixel count of 36.3 million for definition and image quality. Its ultra-high resolution (0.18 cm/pixel) made the identification of rape seedlings plant number possible, although it captured RGB (red, green, blue) images. In fact, there were several color vegetation indices based on visible band. And the research on them had an outstanding achievement. This article compared the ExG, ExG-ExR, normalized green-red difference (NGRD) and green leaf index (GLI), which were commonly used in the study, especially in the UAV digital image system. The result of image segmentation showed that all the color vegetation indices could be suitable for the rape seedlings area identification and extraction. In the end, ExG-ExR was chosen in this paper because it was matched with the area of rape seedlings in RGB image best, while GLI had a lot of noise. Before feature extraction, there were some post-processings for the segmentation objects, such as vectorization, buffer analysis, abnormal object elimination and field survey data input. Finally, in 24 quadrats of this research area, 3565 segmentation objects of rape with 15 shape features were identified and extracted. Eighty percent of them were randomly selected for regression, while the remaining 20% were used for testing. A correlation analysis for 15 shape features was conducted to solve the problem of choosing independent variables preliminarily. Three shape features, i.e. the length-to-width ratio of the minimum bounding rectangle, the distribution density of the pixels and the raster number of perimeter, were chosen for the stepwise multiple regression analysis. For the minimum bounding rectangle of vector segmentation rape objects, the length-to-width ratio was calculated. The distribution density of the pixels describes the distribution of the pixels of a segmentation object in space. The most "dense" shape is a square; the more an object is shaped like a filament, the lower its density. The raster number of perimeter is the sum of the raster numbers of border length for a segmentation object. The result indicated that there was a linear relationship between the plant number and 3 selected shape features. The model showed a determination coefficient R2 of 0.803 with high significance, and its standard error of estimate (SEE) was 0.699. Furthermore, strong correlation existed between the ground-measured and model-predicted plant number (R2=0.809) in the test, and the root-mean-square error (RMSE) was 0.728. Overall, by calculating the color vegetation index and analyzing the correlation between rape seedlings plant number and features, the application of low altitude UAV remote sensing system integrated with ultra-high resolution sensor can effectively identify the plant number of rape seedlings planted by mechanical planter. Based on the automatic identification of rape seedlings plant number, the estimation of rape seedlings emergence and the distribution characteristic of row and plant space would be the next study direction for the evaluation on effect and quality of mechanical planting of rape seedlings.%植株数量识别是油菜机械直播效果和质量评估的关键.该文针对油菜机械直播田间植株数量检测中人工统计耗时、费工、效率低下的现实,通过自主搭建的低空无人机遥感平台采集油菜机械直播区域的遥感影像,基于超高分辨率(0.18 cm/pixel)遥感影像计算的颜色植被指数进行油菜目标识别及其形态特征信息提取.结合田间调查数据,采用逐步回归分析方法,建立了机械直播油菜在苗期的株数与遥感特征信息之间的关系.结果显示,油菜目标的株数与其外接矩形的长宽比、像素分布密度和周长栅格数具有较好的线性关系,回归模型的决定系数R2为0.803,并通过显著性检验,其标准估计误差为0.699.模型检验结果显示,观测值与预测值之间的R2为0.809,均方根误差RMSE为0.728.研究结果表明,利用集成超高分辨率传感器的低空无人机遥感平台,通过计算颜色植被指数并分析油菜目标数量与形态特征的相关性,能有效识别油菜机械直播的出苗株数,可为后续油菜机械直播效果的快速、准确评估提供技术支持.