基于多航高多光谱无人机的北方生态交错带土壤含水量反演研究

Retrieval of soil moisture content based on multi-altitude and multispectral UAV data in the northern ecotone

关键词:  无人机  土壤含水量  多光谱  随机森林  XGBoost

Key words:  UAV  soil moisture content (SMC)  multispectral  random forest (RF)  XGBoost

摘要:

土壤含水量（SMC）是调控全球水循环与能量平衡的核心枢纽，深刻影响着气候格局与生态系统的碳循环过程。研究通过设置不同的无人机（UAV）航高（50 m、100 m和150 m），同步测量吉林省白城市洮北区青山镇不同深度（0—10 cm和10—20 cm）的土壤样本，以探究其对构建生态交错带草地土壤含水量多光谱反演模型的影响；利用Green、Red、RedEdge和NIR这4个波段构建植被指数，采用Pearson相关分析筛选敏感变量，并基于随机森林（RF）和极端梯度提升算法（XGBoost）构建反演模型。结果表明：（1）航高对模型精度具有显著影响，50 m高度采集的光谱数据构建的模型性能最优，其反演结果与实测值吻合度最高；（2）土层深度是影响反演效果的关键因素，0—10 cm表层土壤的反演精度显著优于10—20 cm深层土壤；（3）相较于XGBoost算法，基于RF所构建的模型在多数情形下反演精度较高，而XGBoost在10—20 cm深层偶有优势，其中，0—10 cm土层在50 m、100 m和150 m飞行高度下的RF模型验证集决定系数（R²）分别为0.79、0.80和0.63，均方根误差（RMSE）分别为0.013、0.015和0.022；10—20 cm土层RF模型对应指标分别为R²=0.69、0.65、0.43，RMSE=0.019、0.016、0.024。综合比较表明，50 m航高结合RF模型适用于北方生态交错带草地表层土壤水分的快速监测，可为草地生态管理与精准灌溉提供技术支撑。

Abstract:

Background, aim, and scope Soil moisture content (SMC) is a key variable regulating the terrestrial water cycle, surface energy exchange, and carbon balance. Accurate monitoring of SMC is essential for understanding soil-vegetation interactions and supporting precision agriculture and ecological restoration. However, satellite remote sensing suffers from coarse spatial resolution and cannot capture small-scale heterogeneity. Unmanned aerial vehicle (UAV)-based multispectral remote sensing provides flexible, high-resolution observations that can bridge the scale gap. This study aims to assess the effects of UAV flight altitude and soil depth on the accuracy of multispectral SMC inversion models in a typical northern grassland ecotone and to establish an efficient quantitative retrieval framework. Materials and methods The research was conducted in Qingshan Town, Baicheng City, Jilin Province. UAV multispectral imagery was collected at altitudes of 50 m, 100 m, and 150 m, with resolution of 2.31 cm, 4.61 cm and 6.92 cm using a DJI Mavic 3M carrying Green, Red, RedEdge, and NIR sensors. Soil sampling was performed at depths of 0—10 cm and 10—20 cm. A total of 36 sampling points were selected using RTK positioning, and soil moisture was determined by the oven-drying method. Vegetation indices (NDVI, NDRE, GNDVI, LCI, CIG, SR, GRVI, etc.) were calculated from spectral reflectance, and Pearson correlation analysis was used to select indices with |r| ≥ 0.5 as input variables. Random forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms were employed to construct inversion models, and their accuracy was evaluated using R² and RMSE. Results For the RF model, the validation-set R² values for the 0—10 cm soil layer at flight altitudes of 50 m, 100 m, and 150 m were 0.79, 0.80, and 0.63, respectively, with corresponding RMSE values of 0.013, 0.015, and 0.022. For the 10—20 cm layer, the corresponding R² values were 0.69, 0.65, and 0.43, and the RMSE values were 0.019, 0.016, and 0.024. RF achieved higher validation-set R² for the 0—10 cm layer than for the 10—20 cm layer at each altitude. For the 0—10 cm layer, both 50 m and 100 m provided favorable retrieval performance: 50 m yielded a lower RMSE, whereas 100 m yielded a slightly higher R². At 150 m, RF validation-set R² was lower for both soil layers than at 50 m and 100 m. XGBoost achieved a higher validation-set R² than RF for the 10—20 cm layer at 50 m (0.80 versus 0.69). Spatial predictions revealed clear heterogeneity in SMC distribution, showing a gradual increase in moisture from northwest to southeast. The shallow layer showed greater spatial variation, while the deeper layer exhibited more stability but weaker spectral sensitivity. Discussion The decline in model accuracy with increasing flight altitude can be attributed mainly to the combined effects of spatial-resolution loss and mixed pixels. At lower altitude, the UAV imagery retains more detailed information on canopy structure, bare soil patches, and shadow distribution, so vegetation indices respond more sensitively to near-surface moisture variation. As altitude increases, the ground sampling distance becomes larger and each pixel is more likely to integrate multiple surface components, weakening the spectral contrast associated with SMC. Soil depth further modifies this response. The 0—10 cm layer is closer to the soil-atmosphere interface and is more directly linked with surface reflectance, evaporation, and short-term water redistribution, so its moisture condition can be captured more effectively by multispectral imagery. In contrast, the 10—20 cm layer is controlled more strongly by soil texture, root distribution, infiltration, and subsurface storage, and its optical signal is indirect and easily masked by vegetation cover. The algorithm comparison also indicates that RF is more suitable for the small-sample and high-heterogeneity conditions of this study because bagging and random feature selection reduce variance and improve robustness. XGBoost can fit complex nonlinear relationships, but its sequential boosting mechanism is more sensitive to noise, parameter settings, and reduced image quality at higher flight altitudes. These mechanisms explain why RF showed better stability across altitudes and soil depths, whereas XGBoost only showed local advantages under some high-quality data conditions. Conclusions The 0—10 cm soil layer at 50 m altitude provided the optimal inversion results, and the RF model outperformed XGBoost in stability and overall accuracy. UAV-based multispectral data effectively bridge the gap between satellite and field-scale SMC observations, enabling rapid and fine-scale monitoring in grassland ecosystems. Recommendations and perspectives Future studies should expand sampling density and incorporate temporal data to analyze seasonal SMC variations. Integrating UAV multispectral imagery with hyperspectral or microwave data and advanced deep learning algorithms will further enhance subsurface SMC retrieval and promote applications in precision agriculture and ecological management.