Determining Proper Sampling Density and Strategy for Soil Salinity Assessment before Sampling

Evaluating sampling accuracy is of great importance prior to sampling practice. Since it is expected to be a relationship between soil properties and ancillary environmental variables, proper sampling points are those that better simulate the probability distributions of these variables. The possibility of sampling from low probability tails is very low for random and regular network methods, while the Latin Hypercube Sampling (LHS) method divides the input probability distribution into equal intervals and one sample is taken from each interval.  The satellite images band and digital elevation model (DEM) were used as inputs to determine the sampling points in LHS. Three methods including random, regular network and LHS sampling strategies were employed in three sampling densities of 100, 50 and 25 points in the study area of 52,000 ha to simulate the probability distribution of Landsat ETM+ bands and DEM. The simulated probability distributions were then compared with the population distribution and soil salinity data. Results indicated that there is no significant difference between the sampling strategies in 100 and 50 sampling densities to simulate the mean of population; as the sampling density decreased, the Residual Mean Square Error (RMSE) increased. In 25 point sampling density, the RMSE for LHS, regular network, and random sampling methods were 4, 7.83, and 17.5 percents, respectively. Results further indicated that LHS can efficiently simulate the input probability distribution even in low sampling densities. The RMSE values for estimating soil salinity in the original population were 4.22 and 8.66 dS/m in 100 and 25 sampling densities, respectively. Although the calculated RMSE was larger than that of input parameters of image bands and DEM, it could simulate the population probability distribution of soil salinity with enough accuracy. In conclusion, the best sampling method in this research was Latin hypercube method with a density of 25 sampling points.