multivariate adaptive regression splines

Calculation and estimation of evapotranspiration is one of the most important parameters of water management in agricultural engineering projects. The aim of this study was to evaluate the models of gene expression programming (GEP), group method of data handling (GMDH), and Multivariate adaptive regression spline (MARS) to estimating daily reference evapotranspiration. For this purpose, daily data recorded during the last 25-year period (1993-2017) of Aligudarz region (located in the east of Lorestan province) were used. 80% of the data were used for training and the remaining 20% for testing the models. The modeling results showed that only with the maximum temperature and average wind speed can evapotranspiration be estimated with very good accuracy. The error indices of GEP model in testing stage are , the error indices of MARS and GMDH models are .  Comparing the performance of the models showed that the March model performed better than the other models.

Dew point temperature is the temperature to which under constant pressure, air becomes saturated with water vapor. The goal of the present research is to evaluate the capability of Artificial Neural Networks (ANN (and Multivariate Adaptive Regression Splines (MARS) for estimating the dew point temperature using meteorological parameters in Khoy synoptic station located in northwest of Iran. Used meteorological data were including maximum air temperature (Tmax), minimum air temperature (Tmin), mean air temperature (T), relative humidity (RH), maximum relative humidity (RHmax), minimum relative humidity (RHmin), solar radiation (S), wind speed (W), station atmospheric pressure (Pa), actual vapor pressure (ea) and saturate vapor pressure (es). The mentioned parameters were entered to the used models with various combinations as inputs. To assess the models outputs results, root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R2) were employed. On the basis of the obtained results, the actual vapor pressure (ea) and minimum temperature (Tmin) were the most effective parameters in estimating dew point temperature. Also, the results showed that two used models have adequate accurate to estimate dew point temperature using meteorological parameters. However, the MARS had better performance than ANN in estimating dew point temperature. In general, among the used models and parameters, the MARS with single input of the actual vapor pressure and RMSE= 0.343ºcMAE= 0.480 ºcو R2 =0.991, results the best estimation for of dew point temperature in the test state.

Information on aboveground carbon (AGC) is important for managing forests at local level, land management at regional levels, and carbon emissions reporting at national and international levels; therefore, there is a critical need for low-cost and time-saving approaches for quantifying of AGC. According to this, the estimation of aboveground carbon stock from remotely-sensed data has currently attracted a lot of attention. We assessed the capability of Landsat 8-derived spectral variables for AGC estimates in Zagros coppice oak forests by four non-parametric modeling including: random forest (RF), Cubist, Multivariate adaptive regression spline (MARS) and artificial neural networks (ANNs) The study was carried out in part of Zagros forest, in Kermanshah Province. The values of aboveground carbon (AGC) terrestrial references was determined using field measurement data collected in two sites, Gahvareh (very low degraded (LD) site) and SarfiruzAbad (highly degraded (HD) site). Totally, 124 plots (30×30 meters) surveyed in two studied sites by the systematically-gridded plot design and AGC was calculated by developed species-specific allometric equations for Brant oak trees. For modeling AGC using the remotely-sensed data, we used different Landsat 8 derived variables such as single raw bands, simple band ratios, vegetation indices, principle component analysis and tasseled cap as independent variables and calculated AGC values in plots as dependent variable. The assessment of accuracies of four used non-parametric modeling RF, Cubist, MARS and ANNs and was evaluated by “Leave-one-out” cross validation via criteria such as coefficient of variation (R2), root mean square error (RMSE) and bias. The results showed the accuracy of AGC estimates in LD site was better than HD site. The comparison of used modeling methods revealed that there are not significant difference in performances and accuracies of used non-parametric approaches. In additional, using the total plots from two test sites in one model caused the increase the results for HD site estimates. Results showed R2 and relative RMSE values of approximately 0.7 and 32% cross-validated (combined two studied sites) for modeling AGC using Landsat derived variables, which reveals the high potential of Landsat 8 images and non-parametric modeling methods for quantifying AGC in cost and time saving approach for Zagros forests.