微分光谱连续小波系数估测雅氏落叶松尺蠖危害下的落叶松失叶率

害虫引起的林木失叶会严重威胁森林健康。森林虫害遥感监测与评价中快速、准确获取失叶信息十分重要。基于此，针对雅氏落叶松尺蠖引起的落叶松失叶灾象，在蒙古国开展受害林木光谱测量和失叶率估测试验。首先通过光谱实测数据的处理，得到微分光谱反射率（DSR，对光谱反射率求一阶导数）和微分光谱连续小波系数（DSR-CWC，利用Biorthogonal，Coiflets，Daubechies和Symlets等4种小波系的36个母小波基函数对DSR进行连续小波变换），分析DSR和DSR-CWC对失叶率的敏感性，进而借助MATLAB的Findpeaks（Fp）函数自动寻找DSR和DSR-CWC的敏感波段并确定其对应的敏感特征，然后利用连续投影算法（SPA）对敏感特征进行降维处理，最后利用敏感特征建立偏最小二乘回归（PLSR）和支持向量机回归（SVMR）失叶率估测模型，并与逐步多元线性回归（SMLR）模型进行比较。研究结果表明：①DSR-CWC与DSR相比，对失叶率变化的敏感性更显著且敏感波段亦较多，其敏感波段主要分布于三个吸收谷（440～515，630～760和1 420～1 470 nm）和三个反射峰（516～620，761～1 000和1 548～1 610 nm）范围内。说明DSR-CWC能够增强光谱反射和吸收特征。②Fp与SPA结合模式（Fp-SPA）不仅能够快速、客观选择敏感特征，而且对特征有效降维，是一种光谱敏感特征选择的有效方法。③4种小波系的最优母小波基分别为bior2.4，coif2，db1和sym6，其中db1的失叶率估测性能最稳定，精度最高。④对DSR进行连续小波变换能够提高失叶率估测精度，在DSR-CWC中db1-PLSR模型（R2_M=0.934 0，RMSE_M=0.089 0）提高的最为显著，比DSR-PLSR的R2_M提高了0.047 5并且比DSR-PLSR的RMSE_M降低了0.024 9。⑤利用DSR-CWC建立的PLSR和SVMR模型估测精度类似，其精度优于SMLR模型。可见，DSR-CWC比DSR失叶率估测更有潜力，可为森林虫害遥感监测中提供重要参考。

Estimation of Leaf Loss Rate in Larch Infested with Erannis Jacobsoni Djak Based on Differential Spectral Continuous Wavelet Coefficient

Defoliation caused by insect pests severely threatens the health and safety of forests; the rapid and accurate acquisition of information regarding leaf loss is of considerable significance to the remote sensing monitoring and estimation of forest pests. Based on this, we conducted spectral measurements of infested trees and tested leaf loss rate estimation owing to larch defoliation caused by Erannis jacobsoni Djak in Mongolia. Differential spectral reflectance (DSR, first derivative of spectral reflectance) and continuous wavelet coefficient of differential spectral reflectance (DSR-CWC, continuous wavelet transform of DSR carried out using 36 mother wavelet basis functions of four wavelet families: biorthogonal, coiflets, daubechies and symlets) were obtained based on the processing of spectral measurement data. The sensitivity of DSR and DSR-CWC with respect to the estimation of leaf loss rate was analyzed, following which the sensitive bands of DSR and DSR-CWC were automatically identified using the Findpeaks (Fp) function of MATLAB and the sensitive features identified. Dimension reduction of the sensitive features was processed using a successive projections algorithm (SPA). Partial least squares regression (PLSR) and support vector machine regression (SVMR) models for estimating leaf loss rate were established based on these sensitive features and their effectiveness was compared with that of stepwise multiple linear regression (SMLR) models. The results showed that: ①DSR-CWC was determined to be more sensitive than DSR to changes in leaf loss rate in infested larch, with more sensitive bands, mainly distributed in three absorption valleys (440～515, 630～760 and 1 420～1 470 nm) and three reflection peaks (516～620, 761～1 000 and 1 548～1 610 nm). This finding reflects the fact that DSR-CWC can enhance spectral reflection and absorption characteristics. ②The use of the combination pattern of Fp and SPA (Fp-SPA) was an effective method for the selection of sensitive spectral features that could not only select these features quickly and objectively but also effectively reduce dimensions. ③The optimal mother wavelet bases for the four wavelet families respectivelywere bior2.4, coif2, db1, and sym6; db1 had the most stable performance and accuracy for leaf loss rate estimation. ④The continuous wavelet transform of DSR could improve the accuracy of leaf loss estimation; db1-PLSR (R2_M=0.934 0, RMSE_M=0.089 0) exhibited the most obvious improvement, achieving an R2_M that was 0.047 5 higher than that of DSR-PLSR and an RMSE_M that was 0.024 9 lower than that of DSR-PLSR. ⑤The estimation accuracy of the PLSR and SVMR modelsestablished based on DSR-CWC was either similar to or better than that of the SMLR models. DSR-CWC thus estimated leaf loss rate more effectively than DSR did. It can be seen that DSR-CWC has more potential than DSR in estimating leaf loss rate, and it can provide important reference for remote sensing monitoring of forest pests.