基于无人机遥感影像的玉米苗期株数信息提取

准确、快速地获取玉米苗期株数对于育种早期决策起着至关重要的作用。该文利用2017年6月于北京市小汤山镇采集的无人机影像,首先对比分析RGB、HSV、YCbCr及L*A*B 4种色彩空间,变换优选HSV颜色模型对无人机影像前景(作物)与后景(土壤背景)进行分割,得到分类二值图。然后利用骨架提取算法及多次去毛刺处理等数学形态学流程提取玉米苗形态,得到高精度作物形态骨架,结合影像尺度变换剔除噪声影像,将影像分为多叶、少叶2类,经Harris、Moravec和Fast角点检测识别结果对比,Harris角点检测算法可以较好地提取玉米苗期影像的株数信息。结果表明,少叶类型识别率达到96.3%,多叶类型识别率达到99%,总体识别率为97.8%,将目前传统影像识别精度提高了约3%。同时在多个植株叶片交叉重叠覆盖的情况下,该文的研究方法有良好的适用性。通过无人机影像提取玉米苗期作物准确数目是可行的。该文采用了数学形态学的原理,通过HSV色彩空间变换得到的二值图,从无人机影像中识别提取玉米苗期形态信息,利用影像尺度缩放变换去除噪点,优化骨架识别算法使得识别精度大大提高,最后采用角点检测从无人机影像中直接读取玉米材料小区内的具体数目,该方法节省了人力物力,为田间大面积测定出苗率及最终估产提供了参考。

Extraction of maize seedling number information based on UAV imagery

Accurate and rapid acquisition of maize seedling number plays an important role in early decision-making of breeding. The aim of this work was to use the UAV images collected in Xiaotang mountain, Beijing in June 2017, to recognize and extract the number of maize at seedling stage by establishing a set of digital morphological process. The UAV was flying at a height of about 40 m, and the image data were collected under clear and wind-free conditions. Four color space transformation models (RGB, HSV, YCbCr and L*A*B) were compared and analyzed. The model which showed minimum noise and maintains more pixels was selected to divide image. The foreground (crop) and background (soil) were separated to obtain binary graph. Based on the image scale transformation principle, the images were divided into two types according to the number of leaves. According to the experimental design requirements, different maize breeding materials were growing naturally under the same water and fertilizer conditions. However, it was difficult to control the growth situation of the material area accurately, so there was obvious growth difference between different varieties. When the UAV digital image data were processed, the number of plants was extracted at the same scale, but the result was not ideal, and the recognition rate was only 90.6%. In order to improve the recognition rate, maize breeding materials were classified into two leaf type and multi leaf type according to different growth potential. The skeleton extraction algorithm and multiple deburring processes were utilized to extract crop shape skeleton with high accuracy. Finally, the corner detection results of Harris, Moravec and Fast were compared regarding recognition rate, the leakage recognition rate, the error recognition rate and the operation efficiency. Finally, the Harris corner detection algorithm was used to better extract the number of the maize seedling. Considering that there would be some overlapping leaves in the images, this study showed a comparison between computer identification and actual plant growth. The results showed that this method was still reliable under leaf overlapping conditions. At the same time, the possible errors were analyzed: 1) the influence of the wind during UAV flight; 2) the impact of weeds in the field; 3) the wrongful identification of a single plant as two, but the overall accuracy was still reliable. Through accuracy verification, the leaf recognition rate reached 96.3%, the multi-leaf recognition rate reached 99%, and the overall recognition rate was 97.8%, which proved that it was feasible and reliable to extract corn seedlings from UAV images. This research adopted the principle of mathematical morphology, obtained the binary image by HSV color space transform, recognized and extracted the maize seedling morphological information of UAV images, removed most of the material in the cell skeleton recognition accuracy optimization noise by using image zoom scale transform. By using image zoom scale transform, the recognition accuracy was greatly improved, directly distinguished the specific number of maize materials within the UAV image. This method saved the manpower and material resources and provided strong support for the field of large area determination of germination rate and final yield.