【摘要】 高效获取作物叶绿素含量对农业生产具有重要意义，可为农户提供精准的数据基础，实施及时有效的田间管理策略，最大程度提高作物生长效率和产量。本文采用无人机多光谱技术，通过连续2年的田间试验(2021—2022年),采集了大豆叶片SPAD值和相应的无人机多光谱图像，建立了冠层纹理特征和随机提取的纹理指数组合。通过分析这些参数与大豆叶片SPAD值的相关性，选取相关系数显著(P<0.05)的参数作为模型输入变量(组合1:纹理特征；组合2:随机提取的纹理指数；组合3:纹理特征和随机提取的纹理指数组合)。随后利用随机森林(Random forest, RF)、反向传播神经网络(Backpropagation neural network, BPNN)和极端梯度提升树(Extreme gradient boosting, XGBoost)3种机器学习模型对大豆叶片SPAD值进行估算建模。结果表明：大部分纹理特征与大豆叶片SPAD值的相关系数达到显著水平(P<0.05),MEA为与大豆叶片SPAD值相关系数最高的纹理特征，其相关系数为0.642。而随机组合的纹理指数与大豆叶片SPAD值相关系数达到极显著水平(P<0.01),差值纹理指数(DTI)为与大豆叶片SPAD值相关系数最高(0.761)的纹理指数，纹理组合为(MEA2,MEA5)。总体而言，输入组合3并结合XGBoost模型对大豆叶片SPAD值进行估算的效果最佳，估算模型验证集决定系数R~2为0.883,均方根误差为0.886,平均相对误差为1.638%。研究结果为无人机多光谱监测大豆叶片叶绿素含量奠定了基础，为快速评估作物生长情况提供了依据。更多还原

【Abstract】 Efficient acquisition of crop chlorophyll content is of great significance to agricultural production, as it can provide farmers with precise data foundation for implementing timely and effective field management strategies, thereby maximizing crop growth efficiency and yield. Unmanned aerial vehicle(UAV) multispectral technology was employed, and through two consecutive years of field experiments(2021—2022), soybean leaf SPAD values and corresponding UAV multispectral images were collected to establish combinations of canopy texture features and randomly extracted texture indices. By analyzing the correlation between these parameters and soybean leaf SPAD values, parameters with significant correlation coefficients(P<0.05) were selected as input variables for the models(Combination 1: texture features; Combination 2: randomly extracted texture indices; Combination 3: a combination of texture features and randomly extracted texture indices). Subsequently, three machine learning models, namely random forest(RF), backpropagation neural network(BPNN), and extreme gradient boosting(XGBoost), were utilized to develop estimation models for soybean leaf SPAD values. The results indicated that the correlation coefficients between most texture features and soybean leaf SPAD values exceeded the significant level(P<0.05). MEA was the texture feature with the highest correlation coefficient(0.642) of soybean leaf SPAD values. In contrast, the correlation coefficients between randomly combined texture indices and soybean leaf SPAD values reached extremely significant level(P<0.01). The difference texture index(DTI) was the texture index with the highest correlation coefficient(0.761) of soybean leaf SPAD values, and its corresponding texture combination was(MEA2, MEA5). Overall, the best estimation performance for soybean leaf SPAD values was achieved when using Combination 3 as the input variable combined with the XGBoost model. For the validation set of the estimation model, the coefficient of determination(R~2) was 0.883, the root mean square error(RMSE) was 0.886, and the mean relative error(MRE) was 1.638%. The findings can lay a foundation for UAV multispectral monitoring of chlorophyll content in soybean leaves and provide a basis for the rapid assessment of crop growth status.更多还原