Remote Sensing of Environment vol:110 issue:1 pages:59-78
in this paper the possibility of predicting salt concentrations in soils from measured reflectance spectra is studied using partial least squares regression (PLSR) and artificial neural network (ANN). Performance of these two adaptive methods has been compared in order to examine linear and non-linear relationship between soil reflectance and salt concentration. Experiment-, field- and image-scale data sets were prepared consisting of soil EC measurements (dependent variable) and their corresponding reflectance spectra (independent variables). For each data set, PLSR and ANN predictive models of soil salinity were developed based on soil reflectance data. The predictive accuracies of PLSR and ANN models were assessed against independent validation data sets not included in the calibration or training phase. The results of PLSR analyses suggest that an accurate to good prediction of EC can be made based on models developed from experiment-scale data (R-2>0.81 and RPD (ratio of prediction to deviation)>2.1) for soil samples salinized by bischofite and epsomite minerals. For field-scale data sets, the PLSR predictive models provided approximate quantitative EC estimations (R-2=0.8 and RPD=2.2) for grids 1 and 6 and poor estimations for grids 2, 3, 4 and 5. The salinity predictions from image-scale data sets by PLSR models were very reliable to good (R-2 between 0.86 and 0.94 and RPD values between 2.6 and 4.1) except for sub-image 2 (R-2=0.61 and RPD=1.2). The ANN models from experiment-scale data set revealed similar network performances for training, validation and test data sets indicating a good network generalization for samples salinized by bischofite and epsomite minerals. The RPD and the R-2 between reference measurements and ANN outputs of theses models suggest an accurate to good prediction of soil salinity (R-2>0.92 and RPD>2.3). For the field-scale data set, prediction accuracy is relatively poor (0.69 > R-2 >0.42). The ANN predictive models estimating soil salinity from image-scale data sets indicate a good prediction (R-2>0.86 and RPD>2.5) except for sub-image 2 (R-2=0.6 and RPD=1.2). The results of this study show that both methods have a great potential for estimating and mapping soil salinity. Performance indexes from both methods suggest large similarity between the two approaches with PLSR advantages. This indicates that the relation between soil salinity and soil reflectance can be approximated by a linear function. (c) 2007 Elsevier Inc. All rights reserved.