Spiking-Hybrid方法与机器学习结合的冬小麦LAI反演

LAI inversion of winter wheat based on Spiking-Hybrid model and machine learning

关键词:  小麦  高光谱  Spiking-Hybrid方法  PROSAIL  少样本

Key words:  wheat  hyperspectral  Spiking-Hybrid method  PROSAIL  few samples

摘要:

【目的】 准确地反演叶面积指数（Leaf Area Index，LAI）对小麦生长诊断和管理调控具有重要意义。目前机器学习方法被广泛应用于作物参数反演，但农业领域原始数据获取成本高，机器学习模型在LAI反演中面临训练数据不足、过拟合等问题。【方法】 文章基于遥感辐射传输PROSAIL模型模拟数据结合实测数据，采用Spiking-Hybrid方法构建混合样本集，再利用随机森林（Random Forest，RF）和偏最小二乘回归（Partial Least Squares Regression，PLSR）机器学习算法来反演叶面积指数。将Spiking-Hybrid方法与数值优化的PROSAIL反演方法、传统混合方法和经验机器学习方法等3种常用的植被性状估计方法进行对比分析。【结果】 Spiking-Hybrid方法在不同小麦生育期的LAI反演中展现出优于其他方法的效果，即使选择不同条数、不同生长地域的实测抽样样本，Spiking-Hybrid方法一直表现出更好的准确度和稳健性。当抽取实测样本为40~60条时模型精度最高，在挑旗期使用60份样本时达到了最佳测试精度（R2=0.85，RMSE=0.78）。【结论】 当实测样本十分有限时，Spiking-Hybrid方法比基于模拟数据的机器学习算法具有更好的反演精度，并且Spiking-Hybrid方法在极少的实测样本量时也能发挥很好的作用。

Abstract:

Purpose Accurately inverting the Leaf Area Index（LAI）is of significant importance for diagnosing wheat growth and managing its regulation. Currently，machine learning methods are widely applied in crop parameter inversion. However，in the agricultural field，the high cost of acquiring raw data leads to challenges such as insufficient training data and overfitting in machine learning models when applied to LAI inversion.Method Based on simulated data from the PROSAIL radiative transfer model combined with measured data，the Spiking-Hybrid method was used to construct a mixed sample set. The Leaf Area Index was then inverted using the Random Forest（RF）and Partial Least Squares Regression（PLSR）machine learning algorithms. The Spiking-Hybrid method was compared and analyzed against three commonly used vegetation trait estimation methods：the numerically optimized PROSAIL inversion method，traditional hybrid methods，and empirical machine learning methods.Result The study showed that the Spiking-Hybrid method exhibited superior performance in LAI inversion across different wheat growth stages compared to other methods. Even when selecting measured sample data of varying sizes and from different growth regions，the Spiking-Hybrid method consistently demonstrated better accuracy and robustness. The model achieved the highest accuracy when 40 to 60 measured samples were used，with the best test accuracy（R2=0.85，RMSE=0.78）obtained during the flagging stage using 60 samples.Conclusion When measured samples are extremely limited，the Spiking-Hybrid method achieves better inversion accuracy than machine learning algorithms based solely on simulated data. Additionally，the Spiking-Hybrid method performs effectively even with a minimal number of measured samples.