基于运动恢复结构算法的油菜NDVI三维分布

植物形态伴随着植物生长过程而发生变化，植物的三维重建对研究植物形态对植物生物量估测、植物病害虫害、基因型表达等有着很重要的意义。目前三维重建方法重建出的三维点云多包含植物的形态、颜色等特征，无法反应植物营养状况（如叶绿素含量）、病虫害胁迫等原因造成有机质空间三维分布改变，同时以往手段都需要专门仪器，携带和作业都受到很大限制。多光谱图像能够反应有机质含量等化学值的分布，在近地面遥感、农产品质量无损检测等发面取得了广泛的应用。该文通过采集31张4叶龄油菜的多光谱图像，使用运动恢复结构算法（structure from motion）方法对其进行空间三维重建，得到油菜的三维点云，并对点云中噪声点进行滤除。以控制点和控制长度对所得模型进行评价，得到长度最大偏差在0.1023 cm，RMSE=0.052599，证明该方法重建所得模型具有较好的空间均匀性与准确性，最后计算NDVI指数空间分布。证明所得模型对将来研究植物营养与病虫害胁迫空间分布有着重要意义。

Rape plant NDVI 3D distribution based on structure from motion

Abstract: Plants' morphology changes in their growing process. There is a great need of plant's morphology study for future use like biomass estimation, illness and insect infestation, genotype and other agricultural applications. For now, 3D reconstruction methods can get plants' morphology information. It is meaningful to integrate organic matte distribution information in 3D model. NDVI has been proved to an important index in remote sensing and has a close relationship with chlorophyll density. In this work, Tetracam ADC multispectral camera was used. It is a broadband multispectral camera which has 3.2 million pixels in Bayer filter layout on CMOS photosensitive unit. Thirty one multispectral images of a rapeseed plant were collected at three different angles under indoor conditions for 3D reconstruction. The rapeseed plant must remain stationary and background must keep unchanged. A chessboard was added to the scene for control length comparison and to increase background texture detail. The photos equally surrounded rapeseed plant and covered every corner of the scene. Computer vision method, i.e. Structure from motion (SFM) was used to process plant's 3D model. Visual SFM was used for 3D reconstruction and the generated dense point cloud contains 120089 3D points. It worked in the following four steps: 1) extraction of SIFT points, an average of 2490 SIFT points of image were found; 2) motion estimation; 3) bundle adjustment, 3D sparse point cloud contained 3345 points with color were built; 4) dense point cloud generation, point cloud contained 120 089 points with R-G-NIR information were built. The point cloud had a lot of outliers, so a statistical outlier removal method was used for filtering. The removed outliers were 2 682 points. Control length from chess board was used to measure 3D model accuracy. The RMSE of spatial uniformity was 0.052599, and the maximum error was 0.1023 cm. The result showed that this 3D model precisely represented rapeseed plant's morphology. The last step was to extract xyz and r-g-nir data from point cloud, to calculate every point's NDVI and to visualize plants' NDVI spatial distribution. The result generated from Visual SFM was a ply format file which contained 10 fields not only xyz-rgb. So six fields of x-y-z-r-g-nir were extracted from original data and NDVI index of every point was calculated. The histogram of rape 3D model's NDVI showed the amount of point distributed on every NDVI value. As the NDVI value of background chessboard paper and desktop were below 0, their NDVI were set 0. To visualize the NDVI spatial distribution, a pseudo-color transform was performed according to color transformation theory. After pseudo color transformation, NDVI values were transformed into RGB color and the result ply file containing six field x-y-z-r-g-b. The results showed that the attempts to integrate multispectral image information into plant 3D reconstruction worked out well and had a potential for plants' organic matters spatial distribution research. Compare to other 3D reconstruction method like structure-light 3D reconstruction and laser scanning, SFM had less limitations including no need for special instrument and accessory; good reconstruction result; and being able to integrate NIR information in the point cloud. In the future, this method can be used for insects and illness positioning, plant stress reaction and some similar study.