基于环境与光谱相似性的橡胶树叶片磷含量局部估测模型

为解决现有基于光谱相似性的局部样本搜索方法存在样本误选引起模型预测性能降低的问题,该研究提出先利用环境因子对叶片样本进行类别划分,然后在与待估测样本具有相同类别的样本集内进行局部样本搜索的方法。为验证该方法的有效性,将其用于实际案例中。在案例研究中,分3个时期（4－6月：抽叶期,7－9月：成熟期,10－12月：衰老期）在相同地块采集橡胶树叶片样品,然后利用该方法分别构建每个时期橡胶树叶片磷含量高光谱估测模型,并将模型预测精度与利用现有局部样本搜索方法构建的模型进行比较。为体现该研究提出方法的稳定性和可靠性,将每个时期采集的叶片样本随机分割5次,然后利用方差分析比较不同模型之间的预测精度是否存在显著差异。结果表明,利用该研究提出的方法构建的3个时期的橡胶树叶片磷含量高光谱估测模型预测精度（抽叶期：RMSE分别为(0.031±0.003)%和(0.030±0.004)%,成熟期：RMSE分别为(0.030±0.002)%和(0.029±0.003)%,衰老期：RMSE分别为(0.026±0.002)%和(0.024±0.003)%）都要高于利用现有局部样本搜索方法构建的高光谱估测模型（抽叶期：RMSE分别为(0.034±0.002)%和(0.034±0.002)%,成熟期：RMSE分别为(0.042±0.002)%和(0.042±0.003)%,衰老期RMSE分别为(0.034±0.003)%和(0.035±0.003)%）,且在成熟期和衰老期的差异达到了P<0.05的显著性水平,这就证明了在进行局部样本搜索时必须要考虑橡胶树叶片样本所处环境的差异,以避免选择到与待估测样本不属于同一环境条件的局部样本,进而可显著提高估测模型的预测性能。

Local model based on environmental similarity and spectral similarity for estimating leaf phosphorus concentration of rubber trees

A local model has been widely used to determine the dynamic relationship between the spectra and leaf phosphorus concentration (LPC) of rubber trees. Some local samples can be assumed as the stationary relationship of LPC-spectra. A similar LPC-spectra can be closely related to the local samples, where the key points can be normally evaluated for the local model. However, the current searching approaches of local samples cannot consider the environmental differences of rubber tree leaf using only spectral similarity. Some leaf samples under the different conditions from the samples to be estimated can be selected to construct the hyperspectral estimation model, resulting in low accuracy of the model prediction. In this study, a new Local Sample Searching using Environmental Similarity and Spectral Similarity (LSS-ESSS) was proposed to evaluate the LPC of rubber trees. Two steps were divided during searching. Specifically, the leaf samples were first classified as different categories, where the environmental factors were taken as group variables. Then, the local sample searching was conducted in the same dataset with the same category as the sample to be estimated. A case study was applied to verify the model in the Hainan Island of China, where there were large areas of rubber tree forests. A field sampling test was conducted three times in the development periods of rubber tree leaf (the period of putting forth buds and leaves from April to June; the period of leaf maturity from July to September; and the period of leaf senescence from October to December). The samples of rubber tree leaf were collected from nine predefined sites in each period. The hyperspectral estimation models in each period were then employed to predict the LPC of rubber trees. The prediction accuracies of the models were compared in the three periods using the local sample searching. The collected leaf samples in each period were randomly divided into the training dataset and test dataset five times, in order to evaluate the stability and reliability of the model. An analysis of variance was then used to determine the significant differences in the prediction accuracy of the models. Results showed that the prediction accuracies of the LSS-ESSS models were much higher than before, indicating the significant differences at P < 0.05 level in the period of leaf maturity and senescence. Consequently, the environmental samples of rubber tree leaves can greatly contribute to improving the prediction performance of the model during local sample searching.