融合无人机多源传感器的马铃薯叶绿素含量估算

叶绿素是衡量作物光合作用的重要指标，监测马铃薯关键生育期叶片叶绿素含量(Leaf chlorophyll content, LCC)至关重要。获取马铃薯块茎形成期、块茎增长期和淀粉积累期的无人机RGB和多光谱影像，提取无人机多光谱影像的光谱反射率构建植被指数(Vegetation index, VIs),利用Gabor滤波器提取RGB影像的纹理信息(Texture information, TIs)。然后利用机器学习SVR-REF方法进行数据降维获取植被指数和纹理特征重要性排序，并采用迭代的方法在植被指数最佳模型中加入纹理信息，观察每次加入的纹理信息对模型的动态影响。最后使用支持向量机(Support vector machine, SVR)和K-最近邻算法(K-nearest neighbor, KNN)2种机器学习方法进行建模。结果表明，马铃薯3个关键生育期，加入纹理特征后的2种模型精度和稳定性均有提高，且SVR模型精度优于KNN。块茎形成期，SVR模型建模R²由0.61提升至0.71,RMSE由0.20 mg/g降为0.17 mg/g,精度提升14.2%,验...

Estimation of Potato Chlorophyll Content Based on UAV Multi-source Sensor

Chlorophyll is an important indicator for measuring crop photosynthesis, and monitoring leaf chlorophyll content (LCC) of potatoes during critical growth stages. UAV RGB and multispectral images were obtained during the potato tuber formation, tuber growth, and starch accumulation periods. Vegetation indices (VIs) were extracted from UAV multispectral images, and texture information (TIs) was extracted from RGB images by using Gabor filters. Then, the SVR-REF method was used for data dimensionality reduction to obtain the importance ranking of vegetation indices and texture features, and an iterative approach was used to add texture information to the best vegetation index model and observe the dynamic effect of each added texture information on the model. Finally, support vector machine (SVR) and K-nearest neighbor (KNN) algorithms were used for modeling. The results showed that the accuracy and stability of the two models were improved after adding texture features during the critical growth stages of potatoes, and the SVR model performed better than the KNN model. During the tuber formation period, the SVR modeling R2 was increased from 0.61 to 0.71, and RMSE was decreased from 0.20mg/g to 0.17mg/g, with an accuracy improvement of 14.2%. The validation R2 was increased from 0.58 to 0.66, and RMSE was decreased from 0.19mg/g to 0.17mg/g, with an accuracy improvement of 10.5%. During the tuber growth period, the SVR modeling R2 was increased from 0.59 to 0.67, and RMSE was decreased from 0.16mg/g to 0.14mg/g, with an accuracy improvement of 13.3%. The validation R2 was increased from 0.71 to 0.79, and RMSE was decreased from 0.15mg/g to 0.13mg/g, with an accuracy improvement of 13.3%. During the starch accumulation period, the SVR modeling R2 was increased from 0.62 to 0.69, and RMSE was decreased from 0.17mg/g to 0.14mg/g, with an accuracy improvement of 17.6%. The validation R2 was increased from 0.47 to 0.63, and RMSE was decreased from 0.17mg/g to 0.14mg/g, with an accuracy improvement of 17.6%. In addition, the number of vegetation indices involved in SVR modeling during the three periods were 19, 16, and 3, respectively, and the number of texture features were 4, 2, and 9, respectively. When vegetation indices were unable to respond adequately to chlorophyll content, more texture information was involved in modeling, and the model accuracy was improved significantly, further demonstrating the importance of texture features in chlorophyll content inversion in potatoes.