考虑时空融合环境因子的土壤含水率机器学习反演模型优化

Optimization of Soil Moisture Machine Learning Inversion Model Considering Spatiotemporal Fusion of Environmental Factors

摘要:

植被指数作为构建土壤含水率反演模型的关键要素之一，主要来源于遥感影像的提取。针对高时空分辨率影像难以获取的缺点，采用对象级处理策略的自适应时空融合模型（OL-STARFM）对研究区遥感影像融合，提取融合后的归一化植被指数（NDVI）、地表温度（LST）和植被干旱指数（TVDI）作为环境变量，结合土地利用类型、土壤质地、蒸散量、高程、坡向、坡度、原始影像植被干旱指数（TVDI）、归一化植被指数（NDVI）、地表温度（LST），以及气温、降水量和风速作为建模因子，构建基于多元线性逐步回归（MLSR）、随机森林（RF）和梯度提升机（GBM）3种方法的土壤含水率反演模型，并进行优化分析。研究结果表明：地表温度是影响土壤含水率空间变异性的关键影响因素（R为-0.46），其次为蒸散量（-0.43）、气温（-0.39）、融合后归一化植被指数（0.38）、原始归一化植被指数（0.36）、土地利用类型（0.31）、融合后干旱植被指数（-0.3）、原始干旱植被指数（-0.28）、降水量（0.27）、土壤质地（0.27）、坡向（-0.25）、高程（0.26）、坡度（-0.20）及风速（-0.20）；MLSR表现出较强的模型线性处理能力。非线性处理中RF回归模型最稳定，GBM模型则具有最高的精确度，R2为0.910，MAE、MSE及RMSE分别为2.12%、6.89%和2.62%；多元逐步回归方法在土壤含水率反演过程中预测准确率较低，显示出线性模型在处理复杂关系处理时的局限性；OL-STARFM融合方法提取的TVDI和NDVI与土壤含水率的相关系数分别为-0.41和0.38，均高于单一影像提取的植被指数与土壤含水率的相关性，并且有效提高了土壤含水率反演模型的精度，表明该方法在土壤含水率反演模型构建中的可行性，为获取连续的高时空分辨率影像进而有效连续监测土壤含水率提供了理论依据。

Abstract:

As one of the key elements to construct soil moisture inversion model, vegetation index mainly comes from the extraction of remote sensing images. In view of the shortcomings that high spatiotemporal resolution images are difficult to obtain, the adaptive spatiotemporal fusion model (OL-STARFM) with object-level processing strategy was used to fuse the remote sensing images in the study area, and the normalized difference vegetation index (NDVI), land surface temperature (LST) and temperature vegetation dryness index (TVDI) were extracted as environmental variables, combined with land use type, soil texture, evapotranspiration, elevation, aspect, slope, original image vegetation drought index (TVDI), normalized vegetation index (NDVI), land surface temperature (LST), as well as temperature, precipitation and wind speed as modeling factors, a soil moisture inversion model based on three methods, namely multiple linear stepwise regression (MLSR), random forest (RF) and gradient booster (GBM), was constructed and optimized. The results showed that land surface temperature was the key influencing factor affecting the spatial variability of soil moisture (R was -0.46), followed by evapotranspiration (-0.43), air temperature (-0.39), F_NDVI (0.38), NDVI (0.36), land use type (0.31), F_TVDI (-0.3), TVDI (-0.28), precipitation (0.27), soil texture (0.27), slope aspect (-0.25), elevation (0.26), slope (-0.20) and wind speed (-0.20). MLSR showed strong model linear processing ability. In the nonlinear processing, the RF regression model was the most stable, and the GBM model had the highest accuracy, with R2 of 0.910, and MAE, MSE and RMSE were 2.12%, 6.89% and 2.62%, respectively. The prediction accuracy of the multiple stepwise regression method in the process of soil moisture inversion was low, which showed the limitations of the linear model in dealing with complex relationships. The correlation coefficients between TVDI and NDVI extracted by the OL-STARFM fusion method and soil moisture were -0.41 and 0.38, respectively, which were higher than the correlation between vegetation index and soil moisture extracted from a single image, and effectively improved the accuracy of the soil moisture inversion model, indicating the feasibility of the method in the construction of soil moisture inversion model, and providing a theoretical basis for obtaining continuous high spatiotemporal resolution images for effective continuous monitoring of soil moisture.