Modelling end‑of‑season soil salinity in irrigated agriculture through multi‑temporal optical remote sensing, environmental parameters, and in situ information

Abstract

Accurate information of soil salinity levels enables for remediation actions in long-term operating irrigation systems with malfunctioning drainage and shallow groundwater (GW), as they are widespread throughout the Aral Sea Basin (ASB). Multi-temporal Landsat 5 data combined with GW levels and potentials, elevation and relative topographic position, and soil (clay content) parameters, were used for modelling bulk electromagnetic induction (EMI) at the end of the irrigation season. Random forest (RF) regressionwas applied to predict in situ observations of 2008–2011 which originated from a cotton research station in Uzbekistan. Validation, i.e. median statistics from 100 RF runs with a holdout of each 20% of the samples, revealed that mono-temporal (R2: 0.1–0.18, RMSE: 16.7–24.9 mSm−1) underperformed multi-temporal RS data (R2: 0.29–0.45; RMSE: 15.1–20.9 mSm−1). Combinations of multi-temporal RS data with environmental parameters achieved highest accuracies (R2: 0.36–0.50, RMSE: 13.2–19.9 mSm−1). Beside RS data recorded at the initial peaks of the major irrigation phases, terrain and GW parameters turned out to be important variables for the model. RF preferred neither raw data nor spectral indices known to be suitable for detecting soil salinity. Unexplained variance components result from missing environmental variables, but also from processes not considered in the data. A calibration of the EMI for electrical conductivity and the standard soil salinity classification returned an overall accuracy of 76–83% for the period 2008–2011. The presented indirect approach together with the in situ calibration of the EMI data can support an accurate mapping of soil salinity at the end of the season, at least in the type of irrigation systems found in the ASB.