Soil salinization is one of the serious environmental problems in arid and semiarid regions. As an effective technique for monitoring soil salinity, remote sensing (RS) technology has been widely used to estimate soil salinity in recent years. Previous studies on soil salinity mapping based on RS images adopted linear regression (LR) between the field measured of electrical conductivity (EC) and the RS data. It is expected that nonlinear regression (NLR) models improve the accuracy of soil salinity mapping over LR. The main objectives of this study are: (1) evaluation the capability of various NLR models for estimating soil salinity based on optical Sentinel-2 RS images, (2) feature selection for soil salinity estimation, and (3) updated and accurate soil salinity map production in the dried lake bed of Urmia Lake. The investigated NLR models include: polynomials, rational functions, powers, exponential, gaussian, logarithmic, and sum of sinusoidal functions with different degrees. All these regression models were calibrated and evaluated separately based on 8 visible and infrared bands of the Sentinel-2 image and 17 salinity indices to estimate soil salinity in the dried lake bed of Urmia Lake (Iran). The evaluation results confirmed the superiority of the NLR models over the LR model for soil salinity estimation. The polynomial degree 3 (Poly-3) based on S3 index ( $${\text{S}}3 = \frac{G \times R}{B}$$ S 3 = G × R B ) could predict EC value with better accuracy than the best LR model (based on narrow NIR band). The R 2 and RMSE of the Poly-3 model were 0.98 and 8.16 dS/m while corresponding values of the best LR model were 0.88 and 20.85 dS/m in test samples, respectively. In general, the results show that the NLR models, along with RS data, have enough accuracy to estimate soil salinity. To compare these methods visually and estimate salt’s distribution and concentration in this area, soil salinity maps were predicted by the best NLR model ( $${\text{EC}} = 1.63 \times 10^{ - 10} \times {\text{S}}3^{3} - 9.95 \times 10^{ - 6} \times {\text{S}}3^{2} + 0.11 \times {\text{S}}3 - 151.7$$ EC = 1.63 × 10 - 10 × S 3 3 - 9.95 × 10 - 6 × S 3 2 + 0.11 × S 3 - 151.7 ) and the other linear and NLR models in the dried lake bed of Urmia Lake.

中文翻译：

---

使用光学 Sentinel-2 图像和非线性回归模型估算乌尔米亚湖干涸湖床中的土壤盐度

土壤盐渍化是干旱半干旱地区严重的环境问题之一。【摘要】：遥感(RS)技术作为一种有效的土壤盐分监测技术,近年来被广泛应用于土壤盐分估计。以前基于 RS 图像的土壤盐分制图研究采用了现场测量的电导率 (EC) 和 RS 数据之间的线性回归 (LR)。预计非线性回归 (NLR) 模型会提高土壤盐度映射的准确性，而不是 LR。本研究的主要目标是：（1）基于光学 Sentinel-2 RS 图像评估各种 NLR 模型估计土壤盐度的能力，（2）土壤盐度估计的特征选择，以及（3）更新和准确的土壤盐度在乌尔米亚湖干涸的湖床上制作地图。研究的 NLR 模型包括：多项式、有理函数、幂、指数、高斯、对数和不同阶正弦函数的和。所有这些回归模型都基于 Sentinel-2 图像的 8 个可见和红外波段和 17 个盐度指数分别进行校准和评估，以估计乌尔米亚湖（伊朗）干湖床中的土壤盐度。评估结果证实了 NLR 模型优于 LR 模型的土壤盐分估计。基于 S3 指数 ( $${\text{S}}3 = \frac{G \times R}{B}$$ S 3 = G × RB ) 的多项式次数 3 (Poly-3) 可以预测 EC 值比最佳 LR 模型（基于窄 NIR 波段）更高的精度。Poly-3 模型的 R 2 和 RMSE 分别为 0.98 和 8.16 dS/m，而最佳 LR 模型的相应值为 0.88 和 20。在测试样品中分别为 85 dS/m。总的来说，结果表明 NLR 模型与 RS 数据一起具有足够的精度来估计土壤盐度。为了直观地比较这些方法并估计该地区盐的分布和浓度，土壤盐度图由最佳 NLR 模型预测（ $${\text{EC}} = 1.63 \times 10^{ - 10} \times {\text {S}}3^{3} - 9.95 \times 10^{ - 6} \times {\text{S}}3^{2} + 0.11 \times {\text{S}}3 - 151.7$$ EC = 1.63 × 10 - 10 × S 3 3 - 9.95 × 10 - 6 × S 3 2 + 0.11 × S 3 - 151.7 ) 以及乌尔米湖干涸湖床中的其他线性和 NLR 模型。