分数阶微分技术在机载高光谱数据估算土壤含水量中的应用

无人机高光谱遥感为精准农业和农业信息化监测提供崭新视角。高光谱传感器具有厘米级空间分辨率和精细的光谱分辨率，可获取高质量的高光谱数据。然而，高光谱数据通常伴随噪声和数据冗余，高光谱信息利用效率低，常规预处理难以满足精准估算的需求。因此，为解决上述现实问题，针对机载高光谱影像的数据挖掘必不可少。利用分数阶微分(FOD)技术逐像元处理机载高光谱数据（步长为0.1）。通过对比FOD技术与整数阶技术对高光谱数据的改善能力，从光谱层面探寻最佳FOD阶数。在梯度提升回归树（GBRT）算法下构建土壤含水量（SMC）估算模型，最终在最佳模型下评估SMC的空间分布。结果表明：FOD技术提高光谱与SMC的相关系数（r_max=0.768），与原始光谱、一阶微分和二阶微分处理后的光谱同SMC相关系数相比，分别提升0.168，0.157和0.158。FOD技术提升模型估算精度的主因是突出有效光谱信息的作用，特别是与水分胁迫密切敏感的叶绿素、植物结构和水分响应波段（430，460，640，660和970 nm）。即使FOD技术取得理想的结果，不同阶数的效果仍有差异。高阶FOD对影像增加了一定噪声，相较于高阶FOD（1＜阶数＜2），低阶FOD（0＜阶数＜1）对相关性的改善更为明显。FOD技术对SMC估算模型的性能有很大提高，在0.4阶模型下取得最优结果（R2_p=0.874，RMSEP=1.458，RPIQ=3.029）。此外，0.1—0.9阶和1.6—1.9阶的SMC估算模型比整数阶模型更优（R2_p提升0.8%～13.8%），但根据模的RPIQ发现，低阶FOD模型在模型的预测能力方面更强。在0.4阶模型下反演农田土壤水分的空间分布表明干旱区农田SMC具有显著的空间异质性。研究结果表明低阶FOD技术有效地实现对高光谱数据挖掘，从而实现农业SMC的精准估算。该研究提出了针对机载高光谱影像处理的新方法，为干旱区精准农业实施和管理提供新的策略。

Application of Fractional Order Differential Technology in the Estimation of Soil Moisture Content Using UAV-Based Hyperspectral Data

UAV-based remote sensing technique provides a new perspective and platform for precision agriculture and agricultural information monitoring. The hyperspectral sensor has centimeter-level spatial and fine spectral resolution, allowing for the acquisition of high-quality hyperspectral data. However, hyperspectral data often bring question on noise, data redundancy and inefficient use of hyperspectral information, whereas conventional preprocessing is difficult to estimate withhigh-precision. Therefore, data mining for UAV-based hyperspectral images is essential to solve the above problems. Here we used fractional order differential (FOD) to process UAV-based hyperspectral data (with a step length of 0.1). The optimal FOD order is explored at the spectral level by comparing the ability of the FOD technique with the integer order technique to improve the hyperspectral data. Soil moisture content (SMC) estimation models were constructed under the Gradient-Boosted Regression Tree (GBRT) algorithm, and the spatial distribution of SMC was finally evaluated under the best model. The results found that the correlation between the spectrum and SMC alsowas increased (absolute maximum correlation coefficient, r_max=0.768). Compared with the original image and processed images via first and second order derivatives, r_max increased by 0.168, 0.157 and 0.158, respectively. The main reason for the FOD technique to enhance the accuracy of model estimation is to highlight the role of effective spectral information, especially chlorophyll, plant structure and water response bands closely sensitive to drought stress. (430，460，640，660 and 970 nm). By comparison, the low-order FOD (order＜1) is more effective in the image quality, correlation and model accuracy than high-order FOD (order>1). The higher order FOD adds a certain amount of noise to the image, though the FOD technology achieves the desired result. Estimated model achieved the best results in the 0.4-order model (R2_p=0.874，RMSEP=1.458，RPIQ=3.029). In addition, the SMC estimation models of 0.1—0.9 order and 1.6—1.9 order outperformed the integer-order models (R2_p improvement of 0.8%～13.8%), but the lower-order FOD models were found to be stronger in terms of model predictive power based on the RPIQ of the models. The spatial distribution of inverse farmland soil moisture under the 0.4 order model indicated significant spatial heterogeneity of farmland SMC in the arid regions. In conclusion, the low-order FOD technique effectively enables the mining of hyperspectral data to accurately estimate agricultural SMC. This study proposes a new approach to airborne hyperspectral image processing that provides a new strategy for precision agriculture implementation and management in arid regions.