辐射传输耦合集成学习的堆叠模型在多源遥感数据融合反演土壤含水率的潜力

Potential of Stacked Model Integrating Radiation Transfer Coupling and Ensemble Learning for Soil Moisture Content Retrieval from Multi-source Remote Sensing Data

摘要:

土壤含水率（SMC）是农业水分管理与干旱监测的关键参数，准确估算SMC对于推动农业可持续发展至关重要。目前，由于经验模型过于依赖统计关系，导致普适性与可移植性较差;而物理模型虽具明确机理，但在复杂环境下常需做简化，因此单一模型估算SMC的精度仍面临挑战，有必要探索融合经验与物理机制的混合建模策略，以提升SMC估算的稳定性与泛化能力。针对这一问题，以新疆生产建设兵团农八师为研究区域，以Sentinel-1/2卫星为数据源，以地面实测高光谱数据为辅助，耦合辐射传输模型与集成学习，本文提出了一种基于土壤辐射传输模型（BSM）的Sentinel-1/2融合反演土壤含水率模型。基于实测土壤高光谱数据率定BSM模型参数范围，结合光谱响应函数和BSM模型模拟Sentinel-2反射率，作用于模型训练，并对实测数据进行验证。将验证结果作为堆叠模型的第2层输入变量，结合Sentinel-1雷达后向散射系数及其纹理特征，使用XGboost、GRBT、AdaBoost估算SMC。结果表明：在BSM模型参数范围的率定中，模型参数lat和lon表现出良好的线性关系，相较于未率定参数的土壤含水率第1层反演模型（R2=-1.143）精度提升至R2=0.368;基于土壤辐射传输模型与Sentinel-1卫星的融合算法反演SMC精度优于机器学习模型，验证集决定系数R2达0.816;融合Sentinel-1雷达后向散射系数及其纹理特征和Sentinel-2多光谱构建的混合模型中，AdaBoost展现出最佳的反演精度，验证集R2达到0.816，验证了经验模型与基于物理的模型在反演SMC过程中的互补潜力。本研究强调了混合方法兼具物理模型的泛化能力与经验模型的灵活性，为研究区域土壤含水率实时监测及土地资源可持续管理提供了新的技术手段。

Abstract:

Soil moisture content (SMC) is a key parameter in agricultural water management and drought monitoring, and accurate estimation of SMC is crucial for promoting sustainable agricultural development. Currently, due to the excessive reliance of empirical models on statistical relationships, their universality and portability are poor. Although physical models have clear mechanisms, they often need to be simplified in complex environments. Therefore, the accuracy of single model estimation of SMC is still facing challenges. It is necessary to explore a hybrid modeling strategy integrating experience and physical mechanisms to improve the stability and generalization of SMC estimation. In response to this issue, taking the Agricultural Eighth Division of Xinjiang Production and Construction Corps as the research area, using Sentinel-1/2 satellite as the data source, and ground measured hyperspectral data as the auxiliary, the radiative transfer model and ensemble learning was coupled, and a Sentinel-1/2 fusion inversion soil moisture content model was proposed based on soil radiative transfer model (BSM). This method firstly calibrated the parameter range of the BSM model based on measured soil hyperspectral data, and then combined the spectral response function and BSM model to simulate Sentinel-2 reflectance, applied it to model training, and verified the measured data. Finally, the validation results would be used as input variables for the second layer of the stacked model, combined with the Sentinel-1 radar backscatter coefficient and its texture features, to estimate SMC using XGboost, GRBT, and AdaBoost. The results showed that in the calibration of the parameter range of the BSM model, the model parameters lat and lon exhibited a good linear relationship (R2=0.545), and the accuracy was improved as R2=0.368 compared with the first layer inversion model of soil moisture content without calibration parameters (R2=-1.143). The fusion algorithm based on soil radiative transfer model and Sentinel-1 satellite had better accuracy in retrieving SMC than the machine learning model, with a validation set determination coefficient R2 of 0.816. In the hybrid model constructed by integrating the backscatter coefficient and texture features of Sentinel-1 radar with Sentinel-2 multispectral, AdaBoost showed the best inversion accuracy, with a validation set R2 of 0.816, verifying the complementary potential of empirical and physics based models in the inversion of SMC. The research result emphasized that hybrid methods combined the generalization ability of physical models with the flexibility of empirical models, providing technological means for real-time monitoring of soil moisture and sustainable management of land resources in the research area.