基于无人机多源遥感数据和机器学习的高通量棉花估产研究

High Throughput Cotton Yield Estimation Based on Multi-source Remote Sensing Data from Unmanned Aerial Vehicles and Machine Learning

摘要:

为综合利用光谱、冠层结构、纹理特征等信息对棉花进行无人机（Unmanned aerial vehicle, UAV）遥感产量估算并系统地分析光谱、冠层结构、纹理特征等信息对估产的贡献程度，本文在构建基于多源UAV数据棉花估产机器学习模型的基础上，进一步确定了估产的最佳生育时期，并对比了多源传感器数据在棉花产量估算中的效果，最后量化了各类输入特征的贡献度。采集棉花冠层RGB（Red green blue）、多光谱（Multispectral, MS）和激光雷达（Light detection and ranging，LiDAR）3种传感器数据，通过对棉花光谱植被指数与产量进行相关性分析，确定了棉花产量估算最佳生育时期，进而构建了基于偏最小二乘法回归（Partial least squares regression，PLSR）、随机森林回归（Random forest regression，RFR）、极致梯度提升（Extreme gradient boost，XGBoost）3种机器学习模型的棉花产量估算方法，并评估了基于2种最常用的传感器（RGB和MS相机）的性能。最终确定了光谱特征、冠层结构、纹理特征这3类特征信息在产量估算中的贡献度。研究结果表明，盛花期是棉花估产的最佳生育时期；基于盛花期的UAV数据，XGBoost模型取得了最高的产量估算精度（R2为0.70，RMSE为611.31 kg/hm2，rRMSE为10.60%），在对比基于RGB和MS图像数据提取的特征时，基于MS图像数据提取的特征建模结果更好，同时将RGB和MS相机2种传感器数据提取的特征作为输入时，模型结果高于单一传感器；使用夏普利加性解释（Shapley additive explanations，SHAP）算法分析了机器学习模型中各个输入特征对于估产的贡献度，发现基于3种传感器的3种特征信息在产量估算方面都具有重要意义，其中，纹理特征与冠层结构在产量估算中展现出了较好的潜力。本研究可为棉花智慧化管理中高通量棉花产量估算提供理论和技术支持。

Abstract:

Aiming to utilize information from spectral data, canopy structure, and texture features for cotton yield estimation through unmanned aerial vehicle (UAV) remote sensing, while systematically analyzing the contribution of these factors to yield estimation, based on the construction of a machine learning model for cotton yield estimation by using multi-source UAV data, the optimal growth stage for yield estimation was further identified and the effectiveness of multi-source sensor data in estimating cotton yield was compared. Finally, the contribution of various input features was quantified. Data were collected from three types of sensors: RGB (red, green, blue), multi-spectral (MS), and light detection and ranging (LiDAR). By conducting a correlation analysis between cotton spectral vegetation indices and yield, the optimal growth stage for cotton yield estimation was determined. Subsequently, yield estimation methods were developed by using three machine learning models: partial least squares regression (PLSR), random forest regression (RFR), and extreme gradient boosting (XGBoost). The performance of models based on the two most commonly used sensors (RGB and MS cameras) was evaluated. The results confirmed that the flowering stage was the optimal growth period for cotton yield estimation. Using UAV data from the flowering stage, the XGBoost model achieved the highest yield estimation accuracy (R2 was 0.70, RMSE was 611.31 kg/hm2, rRMSE was 10.60%). When comparing features extracted from RGB and MS image data, the modeling results based on MS camera data were superior. Additionally, when features extracted from both RGB and MS camera data were used as inputs, the model performance exceeded that of single-sensor data. The Shapley additive explanations (SHAP) algorithm was employed to analyze the contribution of each input feature in the machine learning models for yield estimation. It was found that the three types of feature information derived from the three sensors were all significant for yield estimation, with texture features and canopy structure demonstrating considerable potential in this regard. The research result can provide theoretical and technical support for high-throughput cotton yield estimation in smart cotton management.