Mahboobeh Tayebi; Mahdi Naderi; jahangard mohammadi; Mahdieh Hosseinjani Zadeh
Abstract
Introduction: Soil texture is one of the majorphysical properties of soils thatplays important roles inwater holding capacity, soil fertility, environmental quality and agricultural developments. Measurement of soil texture elements in large scales is time consuming and costly due to the high volume ...
Read More
Introduction: Soil texture is one of the majorphysical properties of soils thatplays important roles inwater holding capacity, soil fertility, environmental quality and agricultural developments. Measurement of soil texture elements in large scales is time consuming and costly due to the high volume of sampling and laboratory analysis. Therefore, assessing and using simple, quick, low-cost and advanced methods such as soil spectroscopy can be useful. The objectives of this study were to examine two statistical models of Partial Least Squares Regression (PLSR) and Principal Component Regression (PCR) to estimate soil texture elements using Visible and Near-Infrared (VNIR) and Short-Wave Infrared (SWIR) reflectance spectroscopy (400-2450nm).
Materials and Methods: A total of 120 composite soil samples (0-10 cm) were collected from the Kafemoor basin (55º 15' - 55º 25' E; 28º 51' - 29º 11' N), Sirjan, Iran. The samples were air dried and passed through a 2 mm sieve and soil texture components were determined by the hydrometer method (Miller and Keeny 1992). Reflectance spectra of all samples were measured using an ASD field-portable spectrometer in the laboratory. Soil samples were divided into two random groups (80% and 20%) for calibration and validation of models. PLSR and PCR models and different pre-processing methods i.e.First (FD) and Second Derivatives (SD), Multiplicative Scatter Correction (MSC) and Standard Normal Variate (SNV) were applied and compared to estimate texture elements. The cross‐validation method was used to evaluate calibration and validation sets in the first part (80%) and coefficient of determination (R2), Root Mean Square Error (RMSE) and Residual Prediction Deviation (RPD) were also calculated. For testing predictive models, the second part of data (20%) was used and R2 and RMSE of predictive accuracy were calculated.
Results and Discussion: The results of applying two statistical models for estimatingLogClay (%) showed that R2of calibration (R2CV) and validation (R2VAL) datasetranged from 0.22 to 0.72 and 0.12 to 0.54, respectively. The lowest RMSE was computed for PLSR model with SD pre-processing. The highest RPD of calibration (RPDCV) and validation (RPDVAL) were obtained for PLSR with SD pre-processing technique which was classified as a very good and good model, respectively. The results indicated possible prediction of soil clay content by using PCR model with SD pre-processing techniques. In addition, the PCR predicted soil texture elements poorly according to RPD values while the PLSR model with SD pre-processing was the best model for predicting soil clay content. The R2CV and R2VAL of PLSR models for LogSilt (%) varied from 0.34 to 0.73 and 0.27 to 0.58, respectively. The RMSECV varied from 0.14 for FD pre-processing to 0.23 for no-preprocessing and the RMSEVAL rangedbetween 0.18 and0.24. The highest RPDCV (2.07) and RPDVAL (1.59) were obtained for PLSR with FD pre-processing which were classified as very good and good models, respectively. The results of PCR model developments for estimating LogSilt (%) indicated that the highest RPDCV and RPDVAL were, respectively, 1.31 and 1.25 for MSC pre-processing techniques which were rated as poor models. On the contrary to PLSR models, PCR models were not reliable for predicting LogSilt (%).Theresultsof PLSR models for estimatingLogSand (%) revealedthat the highest R2CV and R2VAL were 0.56 and 0.47, respectively and the lowest RMSECV and RMSEVAL were 0.14 and 0.16, respectively which were obtained for SD pre-processing. The RPDCV and RPDVAL values for SD pre-processing in PLSR model were 1.59 and 1.39 which were rated as good and poor performance of predictions, respectively. The highest RPDCV and RPDVALfor PCR models were obtained with the MSC pre-processing indicating poor model. Therefore, PLSR model with SD pre-processing techniques was superior model for estimation of LogSand(%).Overall, PLSR model with SD pre-processing techniques performed better in estimatingclay and sand and PLSR model with FD pre-processing gave better estimate of silt content.
Conclusions: Our finding indicated thatclay and silt contentcan be estimated by using electromagnetic spectrum between VNIR-SWIR region. Further, spectroscopy could be considered as a simple, fast and low cost method in predicting soil texture and PLSR model with SD and FD pre-processing seems to be more robust algorithm to estimateLogClay and LogSilt, respectively.
