Comprehensive modeling of monthly mean soil temperature using multivariate adaptive regression splines and support vector machine

Soil temperature (T _s) and its thermal regime are the most important factors in plant growth, biological activities, and water movement in soil. Due to scarcity of the T _s data, estimation of soil temperature is an important issue in different fields of sciences. The main objective of the present study is to investigate the accuracy of multivariate adaptive regression splines (MARS) and support vector machine (SVM) methods for estimating the T _s. For this aim, the monthly mean data of the T _s (at depths of 5, 10, 50, and 100 cm) and meteorological parameters of 30 synoptic stations in Iran were utilized. To develop the MARS and SVM models, various combinations of minimum, maximum, and mean air temperatures (T _min, T _max, T); actual and maximum possible sunshine duration; sunshine duration ratio (n, N, n/N); actual, net, and extraterrestrial solar radiation data (R _s, R _n, R _a); precipitation (P); relative humidity (RH); wind speed at 2 m height (u ₂); and water vapor pressure (V_p) were used as input variables. Three error statistics including root-mean-square-error (RMSE), mean absolute error (MAE), and determination coefficient (R ²) were used to check the performance of MARS and SVM models. The results indicated that the MARS was superior to the SVM at different depths. In the test and validation phases, the most accurate estimations for the MARS were obtained at the depth of 10 cm for T _max, T _min, T inputs (RMSE = 0.71 °C, MAE = 0.54 °C, and R ² = 0.995) and for RH, V _p, P, and u ₂ inputs (RMSE = 0.80 °C, MAE = 0.61 °C, and R ² = 0.996), respectively.     

Utilization of PSO algorithm in estimation of water level change of Lake Beysehir

In this study, unlike backpropagation algorithm which gets local best solutions, the usefulness of particle swarm optimization (PSO) algorithm, a population-based optimization technique with a global search feature, inspired by the behavior of bird flocks, in determination of parameters of support vector machines (SVM) and adaptive network-based fuzzy inference system (ANFIS) methods was investigated. For this purpose, the performances of hybrid PSO-ε support vector regression (PSO-εSVR) and PSO-ANFIS models were studied to estimate water level change of Lake Beysehir in Turkey. The change in water level was also estimated using generalized regression neural network (GRNN) method, an iterative training procedure. Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R ²) were used to compare the obtained results. Efforts were made to estimate water level change (L) using different input combinations of monthly inflow-lost flow (I), precipitation (P), evaporation (E), and outflow (O). According to the obtained results, the other methods except PSO-ANN generally showed significantly similar performances to each other. PSO-εSVR method with the values of minMAE = 0.0052 m, maxMAE = 0.04 m, and medianMAE = 0.0198 m; minRMSE = 0.0070 m, maxRMSE = 0.0518 m, and medianRMSE = 0.0241 m; minR ² = 0.9169, maxR ² = 0.9995, medianR ² = 0.9909 for the I-P-E-O combination in testing period became superior in forecasting water level change of Lake Beysehir than the other methods. PSO-ANN models were the least successful models in all combinations.     

Comparison of artificial intelligence techniques for prediction of soil temperatures in Turkey

Soil temperature is a meteorological data directly affecting the formation and development of plants of all kinds. Soil temperatures are usually estimated with various models including the artificial neural networks (ANNs), adaptive neuro-fuzzy inference system (ANFIS), and multiple linear regression (MLR) models. Soil temperatures along with other climate data are recorded by the Turkish State Meteorological Service (MGM) at specific locations all over Turkey. Soil temperatures are commonly measured at 5-, 10-, 20-, 50-, and 100-cm depths below the soil surface. In this study, the soil temperature data in monthly units measured at 261 stations in Turkey having records of at least 20 years were used to develop relevant models. Different input combinations were tested in the ANN and ANFIS models to estimate soil temperatures, and the best combination of significant explanatory variables turns out to be monthly minimum and maximum air temperatures, calendar month number, depth of soil, and monthly precipitation. Next, three standard error terms (mean absolute error (MAE, °C), root mean squared error (RMSE, °C), and determination coefficient (R ²)) were employed to check the reliability of the test data results obtained through the ANN, ANFIS, and MLR models. ANFIS (RMSE 1.99; MAE 1.09; R ² 0.98) is found to outperform both ANN and MLR (RMSE 5.80, 8.89; MAE 1.89, 2.36; R ² 0.93, 0.91) in estimating soil temperature in Turkey.

     

Evaluation of methods of spatial interpolation for monthly rainfall data over the state of Rio de Janeiro,Brazil

Five deterministic methods of spatial interpolation of monthly rainfall were compared over the state of Rio de Janeiro, southeast Brazil. The methods were the inverse distance weight (IDW), nearest neighbor (NRN), triangulation with linear interpolation (TLI), natural neighbor (NN), and spline tension (SPT). A set of 110 weather stations was used to test the methods. The selection of stations had two criteria: time series longer than 20 years and period of data from 1960 to 2009. The methods were evaluated using cross-validation, linear regression between values observed and interpolated, root mean square error (RMSE), coefficient of determination (r ²), coefficient of variation (CV, %), and the Willmott index of agreement (d). The results from different methods are influenced by the meteorological systems and their seasonality, as well as by the interaction with the topography. The methods presented higher precision (r ²) and accuracy (d, RMSE) during the summer and transition to autumn, in comparison with the winter or spring months. The SPT had the highest precision and accuracy in relation to other methods, in addition to having a good representation of the spatial patterns expected for rainfall over the complex terrain of the state and its high spatial variability.     

面向带电作业的手臂末端输出力评估方法

为了解决带电作业时手臂末端输出力的准确控制,提出一种基于表面肌电信号（sEMG信号）和支持向量机回归（SVR）实现对手臂末端施力的评估方法.通过手握机械手臂末端的手柄,做往复推拉运动,记录此时手柄处的力传感器的数据F,同时利用3组肌电信号传感器同步采集手臂的肌电信号.将肌电信号提取特征后,与力F组合成样本集合S,在样本集合中随机抽取50%的样本数据作为训练集,分别训练BP神经网络、GRNN神经网络以及SVR神经网络.最后用训练好的神经网络对整个样本集中的力F进行预测,并用均方根误差和相关系数评估模型的预测效果.结果显示,SVR神经网络的预测效果较好,其均方根误差为3.074 0,相关系数为0.951 7.     

基于多种机器学习模型的西北地区蒸散发模拟与趋势分析

基于机器学习方法和多源数据构建高精度蒸散发（Evapotranspiration,ET）产品对研究气候变化背景下干旱、半干旱地区陆地水循环变化具有重要意义。本文利用西北地区12个草地通量站点与卫星遥感产品,基于随机森林、极端梯度提升、支持向量回归和人工神经网络4种机器学习方法构建ET估算模型,制作5 km分辨率ET产品,并分析ET的长期变化趋势。交叉验证结果表明,4种模型的均方根误差都低于0.57 mm·d^-1,R²高达0.73～0.88。SHAP （SHapley Additive exPlanation）可解释性分析表明,4种模型均将净辐射、植被和土壤湿度作为ET估算的重要因子,也能刻画出土壤偏干时土壤水分对ET的限制作用,有较好的物理解释性。多模型集合的ET结果相比单一机器学习模型以及现有遥感产品误差分别降低7％～20％和45％～70％。趋势分析结果显示,西北地区非裸地下垫面在2001—2018年间整体呈现ET增加趋势,平均速率为19 mm/（10 a）。在河套平原和内蒙古中部和东北部地区,ET的增长速率超过降水,这可能会进一步加剧这些地区的干旱化。     

Modeling monthly meteorological and agronomic frost days,based on minimum air temperature,in Center-Southern Brazil

Although Brazil is predominantly a tropical country, frosts are observed with relative high frequency in the Center-Southern states of the country, affecting mainly agriculture, forestry, and human activities. Therefore, information about the frost climatology is of high importance for planning of these activities. Based on that, the aims of the present study were to develop monthly meteorological (F _MET) and agronomic (F _AGR) frost day models, based on minimum shelter air temperature (T _MN), in order to characterize the temporal and spatial frost days variability in Center-Southern Brazil. Daily minimum air temperature data from 244 weather stations distributed across the study area were used, being 195 for developing the models and 49 for validating them. Multivariate regression models were obtained to estimate the monthly T _MN, once the frost day models were based on this variable. All T _MN regression models were statistically significant (p < 0.001), presenting adjusted R ² between 0.69 and 0.90. Center-Southern Brazil is mainly hit by frosts from mid-fall (April) to mid-spring (October). The period from November to March is considered as frost-free, being very rare a frost day within that period. Monthly F _MET and F _AGR presented significant sigmoidal relationships with T _MN (p < 0.0001), with adjusted R ² above of 0.82. The residuals of the frost day models were random, which means that the sigmoidal models performed quite well for interpreting the frost day variability throughout the study area. The highlands of Santa Catarina, Rio Grande do Sul, São Paulo, and Minas Gerais had in average more than 25 and 13 frosts per year, respectively, for F _MET and F _AGR. The F _MET and F _AGR maps developed in this study for Center-Southern Brazil is a useful tool for farmers, foresters, and researchers, since they contribute to reduce frost spatial and temporal uncertainty, helping in planning project for strategic purposes. Furthermore, the monthly F _MET and F _AGR maps for this Brazilian region are the first zoning of these variables for the country.     

Utility of coactive neuro-fuzzy inference system for pan evaporation modeling in comparison with multilayer perceptron

Estimation of pan evaporation (E _pan) using black-box models has received a great deal of attention in developing countries where measurements of E _pan are spatially and temporally limited. Multilayer perceptron (MLP) and coactive neuro-fuzzy inference system (CANFIS) models were used to predict daily E _pan for a semi-arid region of Iran. Six MLP and CANFIS models comprising various combinations of daily meteorological parameters were developed. The performances of the models were tested using correlation coefficient (r), root mean square error (RMSE), mean absolute error (MAE) and percentage error of estimate (PE). It was found that the MLP6 model with the Momentum learning algorithm and the Tanh activation function, which requires all input parameters, presented the most accurate E _pan predictions (r?=?0.97, RMSE?=?0.81?mm?day^?1, MAE?=?0.63?mm?day^?1 and PE?=?0.58?%). The results also showed that the most accurate E _pan predictions with a CANFIS model can be achieved with the Takagi–Sugeno–Kang (TSK) fuzzy model and the Gaussian membership function. Overall performances revealed that the MLP method was better suited than CANFIS method for modeling the E _pan process.     

Comparison of artificial neural network and multivariate linear regression methods for estimation of daily soil temperature in an arid region

Soil temperature (T _S) strongly influences a wide range of biotic and abiotic processes. As an alternative to direct measurement, indirect determination of T _S from meteorological parameters has been the focus of attention of environmental researchers. The main purpose of this study was to estimate daily T _S at six depths (5, 10, 20, 30, 50 and 100?cm) by using a multilayer perceptron (MLP) artificial neural network (ANN) model and a multivariate linear regression (MLR) method in an arid region of Iran. Mean daily meteorological parameters including air temperature (T _a), solar radiation (R _S), relative humidity (RH) and precipitation (P) were used as input data to the ANN and MLR models. The model results of the MLR model were compared to those of ANN. The accuracy of the predictions was evaluated by the correlation coefficient (r), the root mean-square error (RMSE) and the mean absolute error (MAE) between the measured and predicted T _S values. The results showed that the ANN method forecasts were superior to the corresponding values obtained by the MLR model. The regression analysis indicated that T _a, RH, R _S and P were reasonably correlated with T _S at various depths, but the most effective parameters influencing T _S at different depths were T _a and RH.     

Cold air outbreaks along a non-frozen sea channel: effects of wind on snow bands

Wintertime cold air outbreaks along a non-frozen sea channel or a long lake can become destructive if the related bands of heavy snowfall hit onto land. The forcing for such bands is studied with a 2D numerical model set across an east–west sea channel at 60^oN (‘Gulf of Finland’), varying the basic geostrophic wind V _g. Without any V _g opposite coastal land breezes emerge with convergence. This results in a quasi-steady rising motion w _max ~ 7.5 cm/s at 600 m in the middle of the gulf, which can force a snow band. During weak V _g, the rising motion is reduced but least so for winds from 60^o to 80^o (~ENE), when modest alongshore bands could exist near the downstream (Estonian) coast. During V _g of 4–6 m/s from any direction, the land breezes and rising motions are reduced more effectively, so snow bands are not expected during moderate basic flow. In contrast, during a strong V _g of 20–25 m/s from 110^o to 120^o (~ESE) the land breeze perturbations are intense with w _max up to 15–18 cm/s. The induced alongshore bands of heavy snowfall are located in these cases at the sea but quite close to the downstream (Finnish) coast. They can suddenly make a landfall if the basic wind turns clockwise.     

Meshless Surface Wind Speed Field Reconstruction Based on Machine Learning

We propose a novel machine learning approach to reconstruct meshless surface wind speed fields, i.e., to reconstruct the surface wind speed at any location, based on meteorological background fields and geographical information. The random forest method is selected to develop the machine learning data reconstruction model (MLDRM-RF) for wind speeds over Beijing from 2015–19. We use temporal, geospatial attribute and meteorological background field features as inputs. The wind speed field can be reconstructed at any station in the region not used in the training process to cross-validate model performance. The evaluation considers the spatial distribution of and seasonal variations in the root mean squared error (RMSE) of the reconstructed wind speed field across Beijing. The average RMSE is 1.09 m s^?1, considerably smaller than the result (1.29 m s^?1) obtained with inverse distance weighting (IDW) interpolation. Finally, we extract the important feature permutations by the method of mean decrease in impurity (MDI) and discuss the reasonableness of the model prediction results. MLDRM-RF is a reasonable approach with excellent potential for the improved reconstruction of historical surface wind speed fields with arbitrary grid resolutions. Such a model is needed in many wind applications, such as wind energy and aviation safety assessments.     

Data selection using support vector regression

Geophysical data sets are growing at an ever-increasing rate, requiring computationally efficient data selection(thinning)methods to preserve essential information. Satellites, such as Wind Sat, provide large data sets for assessing the accuracy and computational efficiency of data selection techniques. A new data thinning technique, based on support vector regression(SVR), is developed and tested. To manage large on-line satellite data streams, observations from Wind Sat are formed into subsets by Voronoi tessellation and then each is thinned by SVR(TSVR). Three experiments are performed. The first confirms the viability of TSVR for a relatively small sample, comparing it to several commonly used data thinning methods(random selection, averaging and Barnes filtering), producing a 10% thinning rate(90% data reduction), low mean absolute errors(MAE) and large correlations with the original data. A second experiment, using a larger dataset, shows TSVR retrievals with MAE < 1 m s-1and correlations 0.98. TSVR was an order of magnitude faster than the commonly used thinning methods. A third experiment applies a two-stage pipeline to TSVR, to accommodate online data. The pipeline subsets reconstruct the wind field with the same accuracy as the second experiment, is an order of magnitude faster than the nonpipeline TSVR. Therefore, pipeline TSVR is two orders of magnitude faster than commonly used thinning methods that ingest the entire data set. This study demonstrates that TSVR pipeline thinning is an accurate and computationally efficient alternative to commonly used data selection techniques.     

Long-term variation of rainfall erosivity in Calabria (Southern Italy)

The changes in rainfall erosivity have been investigated using the rainfall erosivity factor (R) proposed for USLE by Wischmeier and Smith (R _W-S ) and some simplified indexes (the Fournier index modified by Arnoldus, F, a regional index spatial independent, R _Fr , and a regional index spatial dependent, R _Fs ) estimated by indirect approaches. The analysis has been carried out over 48 rainfall stations located in Calabria (Southern Italy) using data collected in the period 1936–2012 and divided in three sub-periods. The series of the erosivity indexes and of some precipitation variables have been analyzed for evidence of trends using standard methods. The simplified indexes suggested a general underestimation of the rainfall erosivity with respect to R _W-S . The mean underestimation ranged between 23 and 54 % for R _Fr and from 10 to 15 % for R _Fs . Both the sign and the magnitude of the trends were different for the different stations depending on the variable and sub-period considered. In general, the erosivity increased during the period 1936–1955 (1^st sub-period) and during the more recent sub-period (1992–2012, 3^rd sub-period), whereas it decreased during 1958–1977 (2^nd sub-period). The evidence of trends was generally higher for R _W-S than for R _Fr and R _Fs . Focusing on the most recent sub-period (3^rd sub-period), all the variables analyzed showed mainly increasing trends but with different magnitude. More particularly, R _W-S showed a mean increment of 29 %; F, R _Fr and R _Fs increased by 11, 15 and 18 %, respectively; the maximum intensity of 0.5-h precipitation increased by 5 %; and the annual precipitation increased by 22 %. Consequently, it remains difficult to define which precipitation variable plays the dominant role in the temporal variation of rainfall erosivity in the region. However, the overall results suggest that the indexes estimated by indirect procedures (F, R _Fr , and R _Fs ) should be used with caution for climate change analysis, despite they are used for practical purposes considering they are based on easily available information.     

The Canadian boreal snow water equivalent band

Abstract

Satellite and conventional snow water equivalent (SWE) dataseis reveal a well‐defined zone of high winter season SWE (>100 mm) that extends across the northern boreal forest of Canada. SWE coefficient of variation (CV) patterns derived from a monthly averaged (1978–2002) passive microwave derived time series show a high degree of interannual variability across open prairie, southern boreal, and open tundra regions of North America while SWE across the northern boreal forest was highly invariant. The potential existence of a consistent SWE zone resistant to interannual climatic variability over the past 25 years is intriguing in the context of the sensitivity of snow cover to climate variability and change. A ground sampling campaign conceived specifically to evaluate SWE distribution across the northern boreal forest was conducted in northern Manitoba during the 2003–04 winter season. Data from this survey confirmed the SWE gradient across the boreal forest, although satellite‐derived retrievals for the tundra were consistently low.

A series of Canadian Regional Climate Model (CRCM) simulations were conducted to identify feedbacks between the atmosphere and land surface for a domain focused on the northern boreal forest. A control simulation produced monthly patterns of SWE distribution that closely matched the passive microwave retrievals. Water budget computations showed the SWE accumulation pattern to be a function of the modelled regional precipitation pattern, and not the result of surface processes such as melt or evaporation/sublimation. Mean monthly patterns of 850‐hPa fronto genesis forcing corresponded closely to the patterns of accumulated SWE suggesting that lower tropospheric frontal activity was responsible for the snowfall events that led directly to the deposition of the northern boreal SWE band. CRCM sensitivity experiments were conducted with perturbed land cover and terrain. Only subtle differences in SWE accumulation and frontogenesis patterns relative to the control run were found when complete grassland cover was prescribed, though removing orography greatly enhanced the magnitude and zonal extent of the SWE band.     

A climate-isotope regression model with seasonally-varying and time-integrated relationships

This study investigates multivariable and multiscalar climate-??¹⁸O relationships, through the use of statistical modeling and simulation. Three simulations, of increasing complexity, are used to generate time series of daily precipitation ??¹⁸O. The first simulation uses a simple local predictor (daily rainfall amount). The second simulation uses the same local predictor plus a larger-scale climate variable (a daily NAO index), and the third simulation uses the same local and non-local predictors, but with varying seasonal effect. Since these simulations all operate at the daily timescale, they can be used to investigate the climate-??¹⁸O patterns that arise at daily-interannual timescales. These simulations show that (1) complex links exist between climate-??¹⁸O relationships at different timescales, (2) the short-timescale relationships that underlie monthly predictor-??¹⁸O relationships can be recovered using only monthly ??¹⁸O and daily predictor variables, (3) a comparison between the simulations and observational data can elucidate the physical processes at work. The regression models developed are then applied to a 2-year dataset of monthly precipitation ??¹⁸O from Dublin and compared with event-scale data from the same site, which illustrates that the methodology works, and that the third regression model explains about 55% of the variance in ??¹⁸O at this site. The methodology introduced here can potentially be applied to historic monthly ??¹⁸O data, to better understand how multiple-integrated influences at short timescales give rise to climate-??¹⁸O patterns at monthly-interannual timescales.     

Modeling daily reference evapotranspiration (ET0) in the north of Algeria using generalized regression neural networks (GRNN) and radial basis function neural networks (RBFNN): a comparative study

Estimation of reference evapotranspiration (ET₀) is needed to support irrigation design and scheduling, and watershed hydrology studies. There are many available methods to estimate evapotranspiration from a water surface, comprising both direct and indirect methods. In the first part of this study, the generalized regression neural networks model (GRNN) and radial basis function neural network (RBFNN) are developed and compared in order to estimate the reference ET₀ for the first time in Algeria. Various daily climatic data, that is, daily mean relative humidity, sunshine duration, maximum, minimum and mean air temperature, and wind speed from Dar El Beida, Algiers, Algeria, are used as inputs to the GRNN and RBFNN models to estimate the ET₀ obtained using the FAO-56 Penman-Monteith equation (PM56). The performances of the models are evaluated using root mean square errors (RMSE), mean absolute error (MAE), Willmott index of agreement (d) and correlation coefficient (CC) statistics. In the second part of the study, the empirical Hargreaves-Samani (HG) and Priestley-Taylor (PT) equations are also considered for the comparison. Based on the comparisons, the GRNN was found to perform better than the RBFNN, Priestley-Taylor and Hargreaves-Samani models. The RBFNN model is ranked as the second best model.     

Revisiting forest impact on atmospheric water vapor transport and precipitation

Using a robust global precipitation database, we analyze coast-to-interior seasonal precipitation distributions over the world’s major forest regions. We find that the active functioning of boreal forests in summer is associated with an intense ocean-to-land moisture transport, which declines in winter when forest functioning is minimal. This seasonal switch manifests itself as a change in the exponential scale length of precipitation distribution, which exceeds 15?×?10³ km in summer but decreases to (3–4)?×?10³ km in winter. In equatorial rainforests, which are photosynthetically active throughout the year, annual precipitation remains approximately constant, while the coefficient of variation of monthly precipitation significantly declines toward the continent interior. Precipitation over forest during the periods of active forest functioning is always higher than over the adjacent ocean. Such precipitation patterns support the biotic pump concept according to which forest cover drives the ocean-to-land atmospheric moisture transport on a continental scale.     

Estimating sunshine duration from other climatic data by artificial neural network for ET0 estimation in an arid environment

Sunshine duration data are desirable for calculating daily solar radiation (R _s) and subsequent reference evapotranspiration (ET₀) using the Penman–Monteith (PM) method. In the absence of measured R _s data, the Ångström equation has been recommended by the Food and Agriculture Organization (FAO) of the United Nations. This equation requires actual sunshine duration that is not commonly observed at many weather stations. This paper examines the potential for the use of artificial neural networks (ANNs) to estimate sunshine duration based on air temperature and humidity data under arid environment. This is important because these data are commonly available parameters. The impact of the estimated sunshine duration on estimation of R _s and ET₀ was also conducted. The four weather stations selected for this study are located in Sistan and Baluchestan Province (southeast of Iran). The study demonstrated that modelling of sunshine duration through the use of ANN technique made acceptable estimates. Models were compared using the determination coefficient (R ²), the root mean square error (RMSE) and the mean bias error (MBE). Average R ², RMSE and MBE for the comparison between measured and estimated sunshine duration were calculated resulting 0.81, 6.3 % and 0.1 %, respectively. Our analyses also demonstrate that the difference between the measured and estimated sunshine duration has less effect on the estimated R _s and ET₀ by using Ångström and FAO-PM equations, respectively.     

Hourly air temperature driven using multi-layer perceptron and radial basis function networks in arid and semi-arid regions

This study employed two artificial neural network (ANN) models, including multi-layer perceptron (MLP) and radial basis function (RBF), as data-driven methods of hourly air temperature at three meteorological stations in Fars province, Iran. MLP was optimized using the Levenberg–Marquardt (MLP_LM) training algorithm with a tangent sigmoid transfer function. Both time series (TS) and randomized (RZ) data were used for training and testing of ANNs. Daily maximum and minimum air temperatures (MM) and antecedent daily maximum and minimum air temperatures (AMM) constituted the input for ANNs. The ANN models were evaluated using the root mean square error (RMSE), the coefficient of determination (R ²) and the mean absolute error. The use of AMM led to a more accurate estimation of hourly temperature compared with the use of MM. The MLP-ANN seemed to have a higher estimation efficiency than the RBF ANN. Furthermore, the ANN testing using randomized data showed more accurate estimation. The RMSE values for MLP with RZ data using daily maximum and minimum air temperatures for testing phase were equal to 1.2°C, 1.8°C, and 1.7°C, respectively, at Arsanjan, Bajgah, and Kooshkak stations. The results of this study showed that hourly air temperature driven using ANNs (proposed models) had less error than the empirical equation.