基于双端自适应校正的无人机LiDAR玉米株高反演方法研究

Dual-end Adaptive Correction Method for Maize Plant Height Inversion Using UAV-LiDAR

摘要:

玉米株高是评估作物生长状况和产量潜力的关键表型参数。 尽管无人机 LiDAR 具有穿透性优势,但在玉米抽雄吐丝期,高郁闭度冠层导致地面点探测受阻,且品种间雄穗形态差异显著,使得传统基于固定百分位的提取方法面临系统性低估和适应性差的双重挑战。 为此,本研究提出了一种融合了点云密度和高程分布偏度的地面冠层双端自适应校正方法。 该方法旨在修正地面高程的系统性偏差,并通过自适应算法优化冠层顶点提取。 基于 161 个玉米测试品种的 LiDAR 数据,对比分析了本方法与传统方法(直接点云法和 CHM 法)的性能。 结果表明:传统方法在实际应用中存在显著的系统性低估,LiDAR 低百分位点(H p2 )受近地表杂草或枯叶影响,使地面基准被高估约 25 cm,进而导致传统方法的均方根误差(RMSE)超过 14 cm,平均偏差(Bias)约 - 15 cm;提出的双端自适应策略有效解决了上述问题,通过引入地面系统偏差修正量(H offset ),基本消除了系统性偏差(Bias 为 - 0. 04 cm), 结合冠层自适应模型,进一步将 RMSE 降至 4. 29 cm,MAE 降至 3. 22 cm(R2 = 0. 959);该方法具有较好的鲁棒性,误差分析显示,自适应方法显著抑制了极端离群点,且在不同飞行高度(20、30、50 m)下均能保持高精度 (RMSE 小于 4. 6 cm)。本研究提出的方法在不依赖额外辅助数据的前提下,实现了高郁闭度冠层下玉米株高的高精度与鲁棒性反演,为高通量表型分析提供了可靠的技术支撑。

Abstract:

Maize plant height (PH)is a critical phenotypic parameter for evaluating crop growth status and yield potential. Although UAV LiDAR possesses penetrative capabilities, during the tasseling and silking stages of maize, high canopy closure obstructs ground point detection. Additionally, significant variations in tassel morphology among different varieties pose dual challenges of systematic underestimation and poor adaptability for traditional extraction methods based on fixed percentiles. To address this, a ground-canopy dual-end adaptive correction method that integrated point cloud density and elevation distribution skewness was proposed. This method aimed to correct the systematic bias in ground elevation and optimize canopy apex extraction through an adaptive algorithm. Using LiDAR point cloud data from 161 maize test varieties, the performance of the proposed method was comparatively analyzed compared with traditional methods (direct point cloud method, CHM method). The results showed that traditional methods suffered from severe systematic underestimation in practical applications. Due to the obstruction of laser pulse by near-ground weeds or dry leaves, the LiDAR low percentile points (e. g. , H p2 )overestimated the ground reference by approximately 25 cm, leading to an estimation error (RMSE)exceeding 14 cm and a mean bias (Bias)of about - 15 cm. The proposed dual-end adaptive strategy effectively resolved these issues. By introducing a ground bias correction (H offset ), the systematic bias was successfully eliminated (Bias = - 0. 04 cm); combined with an adaptive canopy model, the RMSE was further reduced to 4. 29 cm and MAE to 3. 22 cm (R2 = 0. 959). The method demonstrated strong robustness. Error analysis revealed that the adaptive method significantly suppressed extreme outliers and maintains high accuracy (RMSE < 4. 6 cm)across different flight altitudes (20 m, 30 m, 50 m). The method proposed achieved high-precision and robust retrieval of maize plant height under high canopy closure without relying on additional auxiliary data, providing reliable technical support for high- throughput phenotyping.