نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه مهندسی عمران، دانشکده مهندسی عمران، دانشکدگان فنی ، دانشگاه تهران، تهران، ایران
2 گروه مهندسی کشاورزی، دانشکده کشاورزی، دانشگاه صنعتی اصفهان، اصفهان، ایران.
3 گروه مهندسی عمران، دانشکده مهندسی عمران، دانشکدگان فنی دانشگاه تهران. تهران. ایران.
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Soil temperature is an important dynamic parameter that plays a key role in surface hydrological processes. In this study, a combination of spectral analysis and Box–Jenkins time series models was used to improve the prediction of thermal behavior in soil layers at various depths. Two main scenarios were considered for modeling: the first relying on soil temperature data, and the second including meteorological variables as auxiliary inputs. Model performance was evaluated using MAE, RMSE, R², and AIC criteria. After performing various tests, deterministic components in the time series were identified, and their intensity was assessed using different statistical parameters. The statistical analysis indicated that seasonality plays a more significant role than trend in the soil temperature time series. The developed models showed that combining spectral analysis with ARMA and ARIMA structures significantly improves soil temperature prediction accuracy. At the depth of 100 cm, this method achieved better performance with R² = 0.9750, MAE = 0.83, RMSE = 1.06, and AIC = –221.38, compared to multivariate models. Although in some scenarios the inclusion of meteorological variables such as evapotranspiration, wind speed, and solar radiation improved the results, univariate models based on soil temperature data provided more stable performance. Ultimately, this study demonstrated that combining spectral methods with time series models is an effective and reliable approach for predicting soil temperature at various depths.
کلیدواژهها [English]
EXTENDED ABSTRACT
Soil temperature is one of the fundamental parameters in geoenvironmental engineering, geotechnics, hydrology, and agriculture, influencing many physical, chemical, and biological processes in soil. Therefore, predicting soil temperature can be important for water resource decision-makers. Many researchers adopt a variable-based approach to study soil temperature data. In this approach, a large number of independent variables are used to estimate soil temperature. One suitable method to address these challenges is to perform modeling using stochastic models.
In the first step, the soil temperature data at various depths were divided into training and testing sets to enable performance evaluation of the models. The modeling process was then implemented based on two main scenarios, comprising a total of 12 sub-scenarios. The first scenario focused on univariate modeling of soil temperature and included two main sub-scenarios: the first with data preprocessing, and the second without any preprocessing. In the first sub-scenario, after applying preprocessing techniques, soil temperature was modeled using four different time series approaches: ARIMA, ARMA, AR, and MA. For each approach, 100, 100, 10, and 10 models were developed, respectively, by varying the modeling parameters. In the second sub-scenario, soil temperature data were modeled without any preprocessing, solely using classical Box-Jenkins models including ARIMA, ARMA, and SARIMA. Each of these models was independently implemented and evaluated through 120 separate runs. The second scenario was based on multivariate modeling of soil temperature. In this approach, in addition to soil temperature time series data, auxiliary meteorological variables were incorporated as predictors. This scenario consisted of two sub-scenarios (third and fourth). In the third sub-scenario, soil temperature was modeled using only one auxiliary variable as an additional input. In each case, the selected variable was one of the well-known factors influencing soil temperature.
This study aimed to improve the accuracy of soil temperature forecasting at various depths by combining spectral analysis with Box-Jenkins models. Data stationarity was tested using the KPSS and PP tests, which indicated that the data were non-stationary. The autocorrelation function (ACF) analysis revealed a strong seasonal component with a 365-day periodicity across all time series. Two main modeling scenarios were designed to evaluate the effects of preprocessing techniques and meteorological variables. The results showed that univariate models, which relied solely on soil temperature data, outperformed multivariate models that incorporated meteorological variables. The univariate approach demonstrated better stability and prediction accuracy across different depths. Additionally, the models showed that as depth increased, prediction accuracy improved, with the model performing significantly better at 100 cm compared to 5 cm. The accuracy and efficiency of the model increased by approximately 58-59.5% at greater depths, making it more suitable for predicting soil temperature at deeper levels.
Results showed that incorporating spectral analysis significantly improved the performance of univariate time series models compared to seasonal differencing. Deeper soil layers yielded higher prediction accuracy due to their thermal stability. Model evaluation based on MAE, RMSE, R², and AIC confirmed that preprocessing improves performance. Moreover, statistical analysis of deterministic and stochastic components highlighted the dominant role of seasonality over long-term trends in soil temperature variations. While univariate models outperformed multivariate ones overall, using a single relevant meteorological variable (especially evapotranspiration, wind speed, or solar radiation) as an auxiliary input also proved effective. Future studies are encouraged to test these approaches using hybrid models and under broader environmental conditions.
Fereshteh Nourmohammadi Dehbalaei: Writing – original draft, Formal analysis, Conceptualization, Data curation, Methodology, Validation, Writing – review & editing, supervision, and project administration. Abbas Solimanpour: Formal analysis, Conceptualization, Data curation, Writing – review & editing. Seyed Taghi Omid Naeeni: Formal analysis, Conceptualization, Data curation, Methodology, Validation, Writing – review & editing.
Data is available on reasonable request from the authors.
The authors would like to thank all participants of the present study.
The authors avoided data fabrication, falsification, plagiarism, and misconduct.
The authors declare no conflict of interest
)