摘　要：稗草是稻田中一种典型的杂草，会严重影响水稻的生长发育，并最终导致减产。由于其外观与水稻极其相似，难以区分，给治理带来了极大的挑战。基于室内环境下的稗草识别条件理想，难以推广，因此，在复杂稻田环境下对稗草进行识别和制图具有重要的研究价值和意义。首先，利用无人机载高光谱获取低空稻田影像，经拼接校正和SG（Savitzky-Golay）卷积平滑滤波后，采用连续投影算法（SPA）对区分水稻和稗草的敏感波段进行提取，基于全波段和特征波段，使用支持向量机（SVM）、随机森林（RF）、一维卷积神经网络（1DCNN）和三维卷积神经网络（3DCNN）进行建模。结果表明，SPA-3DCNN对水稻（0.942 0）和稗草（0.893 6）的识别效果最好。SPA选择的7个特征波段（482.523 4 、546.541 5 、675.080 6 、709.138 2 、762.043 1 、922.015 7 和944.637 1 nm）对区分稗草和水稻具有重要的参考价值。随后，将模型应用于整个高光谱数据集，生成稗草的空间分布图与密度图。本文成功探索了无人机载高光谱在复杂稻田环境下稗草识别的可行性，并绘制了稗草空间分布图与密度图，为稗草的治理防范提供了有力的数据支撑。
关键词：稻田杂草识别；无人机；高光谱；空间分布；精准除草
中图分类号：P237；S127；TP79                  文献标志码：ADOI：10.3969/j.issn.1007-7146.2024.04.006

Abstract: Barnyard grass is a typical weed in paddy fields that can severely affect the growth and development of rice, ultimately leading to reduced yields. Its appearance is very similar to that of rice, making it difficult to distinguish and posing significant challenges for management. Ideal conditions for barnyard grass identification are achievable in controlled indoor settings but are difficult to replicate in practical applications. For this reason, identifying and mapping barnyard grass in complex paddy field environments holds significant research value and importance. First, hyperspectral images of the paddy fields were captured using UAVs. After image stitching, rectification, and SG (Savitzky-Golay) convolution filtering, a sequential projection algorithm (SPA) was employed to extract sensitive bands for distinguishing rice from barnyard grass. Modeling was performed across the entire spectral range and selected feature bands, employing support vector machines (SVM), random forests (RF), one-dimensional convolutional neural networks (1DCNN), and three-dimensional convolutional neural networks (3DCNN). The results indicated that SPA-3DCNN achieved the best recognition performance for rice (0.942 0) and barnyard grass (0.893 6). The seven feature bands selected by SPA (482.523 4, 546.541 5, 675.080 6, 709.138 2, 762.043 1, 922.015 7, and 944.637 1 nm) were valuable for distinguishing barnyard grass from rice. Subsequently, the model was applied to the entire hyperspectral dataset to generate spatial distribution and density maps of barnyard grass. This study successfully explored the feasibility of UAV-based hyperspectral identification of barnyard grass in complex paddy field environments, providing strong data support for the management and prevention of barnyard grass.
Key words: identification of weeds in rice fields; unmanned aerial vehicle; hyperspectral; space distribution; precision weeding
（Acta Laser Biology Sinica, 2024, 33(4): 335-346）