不同植被覆盖条件下Sentinel-1/2数据融合监测土壤含盐量模型研究

Monitoring of Soil Salt Content during Different Growth Periods of Crops Based on Sentinel-1/2

摘要:

快速准确获取土壤含盐量（SSC）信息对农业可持续发展至关重要。卫星遥感技术凭借大范围同步监测的优势在SSC监测领域被广泛关注，但其监测精度常面临植被覆盖干扰和灌溉事件等多源误差源的挑战。本研究以Sentinel-1/2卫星数据为基础，结合地面实测数据，对不同植被覆盖条件下的土壤含盐量进行监测，以期明确不同植被覆盖对土壤含盐量遥感监测准确性的影响。首先，根据研究区内植被覆盖度、NDVI变化趋势、作物生育期将全植被覆盖划分为3个时期（D1：早期；D2：中期；D3：后期）；其次，分析不同时期变量（植被指数与极化指数）对不同深度土壤含盐量的敏感性，并利用变量投影重要性(VIP)分析算法筛选变量；最后，结合机器学习算法（SVM、RF与ELM模型），生成各时期不同深度土壤含盐量分布图。结果表明：D2时期变量与SSC的相关性最好，D3时期次之，D1时期最低；雷达遥感与光学遥感数据融合有助于监测作物不同时期的土壤含盐量；RF模型为最佳土壤含盐量监测模型，10~20cm土壤含盐量监测精度最高，R2达到0.79，RMSE为1.62g/kg；从空间分布来看，研究区南部土壤盐渍化程度最重，从深度上看，各时期20~40cm的土壤含盐量最高，从时间变化来看，作物生长0~10cm与10~20cm土壤含盐量呈现增加趋势，20~40cm土壤含盐量呈现减少趋势。研究结果可为区域土壤盐渍化的精准监测和防治提供科学依据。

Abstract:

Accurately and rapidly acquiring soil salinity content (SSC) information is crucial for agricultural sustainable development. Satellite remote sensing technology has attracted extensive attention in SSC monitoring due to its advantage of large-scale synchronous monitoring, but its monitoring accuracy often faces challenges from multiple error sources such as vegetation coverage interference and irrigation events. SSC under different vegetation coverage conditions was monitored based on Sentinel-1/2 satellite data combined with ground-measured data, aiming to clarify the impact of different vegetation coverage on the accuracy of SSC remote sensing monitoring. Firstly, the full vegetation coverage period was divided into three stages (D1: early stage;D2: middle stage;D3: late stage) according to vegetation coverage, NDVI variation trends, and crop growth periods. Secondly, the sensitivity of variables (vegetation indices and polarization indices) to SSC at different soil depths was analyzed, and the variable importance in the projection (VIP) analysis algorithm was used for variable screening. Finally, machine learning algorithms (support vector machine (SVM), random forest (RF), and extreme learning machine (ELM) models) were integrated to generate SSC distribution maps for different soil depths in each stage. Results showed that variables had the highest correlation with SSC in D2, followed by D3 and D1. Fusion of radar and optical remote sensing data contributed to SSC monitoring across different crop stages. The RF model proved optimal for SSC monitoring, with the highest accuracy (R2 of 0.79, RMSE of 1.62g/kg) at 10~20cm soil depth. Spatially, the southern part of the study area exhibited the most severe soil salinization. Vertically, SSC was the highest at 20~40cm across all stages. Temporally, SSC in 0~10c and 10~20cm layers was increased with crop growth, while SSC at 20~40cm showed a decreasing trend. These findings provided a scientific basis for precise monitoring and prevention of regional soil salinization.