Utilizing remote sensing data to estimate soil salinity in irrigated agricultural areas of Colorado's South Platte River basin

Abstract

With the global population projected to reach 10 billion by 2050, agricultural output will need to increase by about 50%. This increase must occur with little to no expansion of farmable land, as climate change is expected to reduce the amount of land available for agriculture. Additionally, there is limited availability of water resources to aid in the increasing output, as irrigated agriculture already accounts for 72% of freshwater withdrawals worldwide – 40% of which is unsustainable. Farmers are being asked to produce more with fewer resources, while challenges like soil salinization complicate the process. In the South Platte River Basin (SPRB) of Colorado, farmers are facing increasing soil and water salinity, which leads to a decrease in crop yields. Remote sensing data provide potential for the development of low-cost methods for monitoring changes and effects of salinity. One such method uses satellite remote sensing data to estimate the soil saturation extract electrical conductivity (ECe, dS/m). Current models for predicting ECe using remote sensing data utilize vegetation indexes to relate crop health to soil health. In this study, two previous methods of estimating salinity were tested and revised to calibrate an equation for the estimation of soil salinity in the SPRB. These ECe modeling methods focused on the application of three different vegetation indexes [i.e., the Normalized Difference Vegetation Index (NDVI), the Canopy Response to Salinity Index (CRSI), and the Crop Water Stress Index (CWSI)], plus weather data from agricultural weather stations. When these existing ECe mapping methods failed to produce accurate ECe results in the SPRB, multiple regression analyses were run using remote sensing data, and new explanatory variables, to calibrate a local model. Both single year and multiyear data ECe models were created utilizing three datasets: one dataset included all fields surveyed, one included only the corn fields surveyed, and one included only the non-corn fields surveyed. Using the highest coefficient of determination (R2) statistic, 20 models from each dataset (60 models total) were selected for further testing. The Corrected Akaike Information Criterion (AICc) was calculated for each of these models to identify the best model for each dataset. When the selected models were tested against data collected during surveys, they showed very small biases (Mean Bias Error less than +/- 0.16 dS/m, Normalized Mean Bias Error less than 6%). For the dataset that included all fields surveyed and the dataset that included only the non-corn fields surveyed, even the most promising models (R2 between 0.63 and 0.7) had poor accuracy with a Normalized Root Mean Square Error (NRMSE) of greater than 30% (33.34-40.83%). The dataset with all surveyed fields was calibrated with observed ECe data that ranged from 0.00-13.33 dS/m with an average of 2.78 dS/m, and the most promising model calibrated on this data predicted ECe values between 0.75-4.43 dS/m with an average of 2.67 dS/m. The dataset which included only the non-corn fields was calibrated with observed ECe data that ranged from 0-13.33 dS/m with an average of 2.91 dS/m, and the most promising model calibrated on this data predicted ECe values between 0.76-5.15 dS/m, with an average of 2.78 dS/m. The modeling results from corn field dataset were generally better at estimating soil salinity with R2 values around 0.84 and NRMSEs of around 20%. The corn field data set was calibrated with observed ECe data that ranged from 0.43-6.86 dS/m with an average of 2.40 dS/m, and the most promising model calibrated on this data predicted ECe values between 1.23-4.81 dS/m with an average of 2.40 dS/m. All tested models from all datasets were accurate to around +/- 1 dS/m (Root Mean Square Error) indicating these models may be useful for field-level management decisions. The number of independent explanatory variables included in the 60 best-performing equations ranged from 2 to 8, with 5-variable models being the most common. Every equation incorporated at least one weather-related variable (precipitation or temperature). Although NDVI was not used as a sole predictor, it was included in nearly all of the top equations (58 out of 60), underscoring its importance in estimating soil salinity. CRSI and CWSI were also prevalent, appearing in approximately two-thirds of the best equations (41 and 42 out of 60, respectively). Additionally, 47 of the 60 equations incorporated the crop growth development-phase data for either vegetation indexes or weather variables, suggesting that the crop development stage is an influential factor in salinity prediction models.