基于机载激光雷达的水稻株高检测方法与泛化能力研究

Rice Plant Height Detection Method and Generalization Ability Based on UAV LiDAR

摘要:

株高作为水稻表型分析与生长状态评估的核心参数，其高通量检测对育种与生产意义重大。机载激光雷达凭借其高精度优势，已成为水稻株高检测的研究热点。但现有研究普遍基于单一特征构建线性回归模型，在多品种、多育种材料等复杂应用场景中精度与泛化能力不足。为此，本研究采用多特征融合策略构建水稻株高非线性回归模型，提升估算精度。在广州市增城区与三亚市崖州区分别开展田间试验，采集5个品种及225份水稻育种材料的株高和激光点云时序数据，构建包含高度百分位数、统计参数和冠层剖面面积的多维度特征体系，利用线性与非线性机器学习算法建立了水稻株高的检测模型。试验表明，多特征非线性预测模型的精度高于线性预测模型，其决定系数（R2）最高达到0.969，均方根误差（RMSE）低至4.73cm，相较于单特征线性模型（R2=0.905，RMSE为8.231cm），R2提升7.2%，RMSE降低42.5%；进一步研究泛化性表明，基于广州增城5个品种数据构建的模型泛化能力，显著低于基于三亚崖州225份育种材料构建的模型，证实水稻样本多样性可有效提升模型鲁棒性。该研究为水稻株高高通量表型分析提供了高精度、高适用性的通用技术框架。

Abstract:

Rice plant height is a core parameter for phenotypic analysis and growth state assessment, and its high-throughput detection is of significant importance for breeding and production. UAV LiDAR, with its high-precision advantages, has become a research hotspot for rice plant height detection. However, existing studies generally constructed linear regression models based on a single feature, which led to insufficient accuracy and generalization ability in complex application scenarios such as multi-variety and multi-breeding material analysis. Therefore, a multi-feature fusion strategy was adopted to build a nonlinear regression model for rice plant height, aiming to improve estimation accuracy. Field experiments were conducted in Zengcheng, Guangzhou, and Yazhou, Sanya, where plant height and laser point cloud time-series data were collected for five rice varieties and 225 breeding materials. A multi-dimensional feature system, including height percentiles, statistical parameters, and canopy profile area, was constructed. Linear and nonlinear machine learning algorithms were used to establish the detection model for rice plant height. The results showed that the accuracy of the multi-feature nonlinear prediction model was higher than that of the linear prediction model, with the highest coefficient of determination (R2) reaching 0.969 and the root mean square error (RMSE) as low as 4.73 cm. Compared with the single-feature linear model (R2= 0.905, RMSE was 8.231 cm), the R2 was increased by 7.2% and RMSE was decreased by 42.5%. Further studies on generalization ability indicated that the model built using data from the five varieties in Zengcheng, Guangzhou, showed significantly lower generalization ability compared with the model constructed from the 225 breeding materials in Yazhou, Sanya, confirming that the diversity of rice samples can effectively improve model robustness. The research result can provide a high-precision and highly applicable general technical framework for high-throughput phenotypic analysis of rice plant height.