Surface Temperature Forecasts Research Articles

Global ocean surface and subsurface temperature forecast skill over subseasonal to seasonal timescales

Global ocean surface and subsurface temperature forecast skill over subseasonal to seasonal timescales

Assessment of forest fire severity and land surface temperature using Google Earth Engine: a case study of Gujarat State, India

Forest fires are a recurring issue in many parts of the world, including India. These fires can have various causes, including human activities (such as agricultural burning, campfires, or discarded cigarettes) and natural factors (such as lightning). The present study presents a comprehensive and advanced methodology for assessing wildfire susceptibility by integrating diverse environmental variables and leveraging cutting-edge machine learning techniques across Gujarat State, India. The primary goal of the study is to utilize Google Earth Engine to compare locations in Gujarat, India, before and after forest fires. High-resolution satellite data were used to assess the amount and types of changes caused by forest fires. The present study meticulously analyzes various environmental variables, i.e., slope orientation, elevation, normalized difference vegetation index (NDVI), drainage density, precipitation, and temperature to understand landscape characteristics and assess wildfire susceptibility. In addition, a sophisticated random forest regression model is used to predict land surface temperature based on a set of environmental parameters. The maps that result depict the geographical distribution of normalized burn ratio and difference normalized burn ratio and land surface temperature forecasts, providing valuable insights into spatial patterns and trends. The findings of this work show that an automated temporal analysis utilizing Google Earth Engine may be used successfully over a wide range of land cover types, providing critical data for future monitoring of such threats. The impact of forest fires can be severe, leading to the loss of biodiversity, damage to ecosystems, and threats to human settlements.

Postprocessing of Ensemble Weather Forecasts Using Permutation-Invariant Neural Networks

Abstract Statistical postprocessing is used to translate ensembles of raw numerical weather forecasts into reliable probabilistic forecast distributions. In this study, we examine the use of permutation-invariant neural networks for this task. In contrast to previous approaches, which often operate on ensemble summary statistics and dismiss details of the ensemble distribution, we propose networks that treat forecast ensembles as a set of unordered member forecasts and learn link functions that are by design invariant to permutations of the member ordering. We evaluate the quality of the obtained forecast distributions in terms of calibration and sharpness and compare the models against classical and neural network–based benchmark methods. In case studies addressing the postprocessing of surface temperature and wind gust forecasts, we demonstrate state-of-the-art prediction quality. To deepen the understanding of the learned inference process, we further propose a permutation-based importance analysis for ensemble-valued predictors, which highlights specific aspects of the ensemble forecast that are considered important by the trained postprocessing models. Our results suggest that most of the relevant information is contained in a few ensemble-internal degrees of freedom, which may impact the design of future ensemble forecasting and postprocessing systems.

Urban land surface temperature forecasting: a data-driven approach using regression and neural network models

The insinuations of the ailments associated with the unrestrained and disorganized proliferation of artificial impervious materials over natural surfaces are prevalent among city dwellers. These impacts can be comprehended by estimating land surface temperature (LST), as it is vital for evaluating urban climate, particularly to explain the intensity of urban heat islands and to define the health and welfare of the planet as well as the living beings. Urbanization-driven landscape changes severely disrupt comfortable living in almost every city, necessitating monitoring and modelling historical, current, and likely future LSTs. This research article proposes two forecasting techniques: Multiple Linear Regression (MLR) and Artificial Neural Network (ANN) models. These models have been widely accepted for the efficient prediction of climatic parameters, including LST, over an urban area. The landscape, elevation, and LST trend served as input to the models for an accurate prediction of LST. The analysis was performed over the Kolkata Metropolitan Area (KMA) with an additional 10 km buffer to understand urban growth and its effect on the LST of the entire region. The two developed models (MLR and ANN) effectively anticipated the LST over the KMA region. A continual increment in the surface temperatures ranging from 1 °C to 4 °C, over existing and likely-predicted urban areas was comprehended. It was anticipated that the regions near the urban areas will also experience severe discomfort and heat waves without proper mitigation measures. This scientific literature provides essential insights for decision-makers, stakeholders, and government officials to articulate new policies and modify the existing ones to create a sustainable and livable urban environment for the inhabitants.

Correcting Nonstationary Sea Surface Temperature Bias in NCEP CFSv2 Using Ensemble-Based Neural Networks

Abstract Sea surface temperature (SST) forecast products from the NCEP Climate Forecast System (CFSv2) that are widely used in climate research and prediction have nonstationary bias. In this study, we develop single- (ANN1) and three-hidden-layer (ANN3) neural networks and examine their ability to correct the SST bias in the NCEP CFSv2 extended seasonal forecast starting from July in the extratropical Northern Hemisphere. Our results show that the ensemble-based ANN1 and ANN3 can reduce the uncertainty associated with parameters assigned initially and dependence on random sampling. Overall, ANN1 reduces RMSE of the CFSv2 forecast SST substantially by 0.35°C (0.34°C) for the testing (training) data and ANN3 further reduces RMSE relatively by 0.49°C (0.47°C). Both the ensemble-based ANN1 and ANN3 can significantly reduce the spatially and temporally varying bias of the CFSv2 forecast SST in the Pacific and Atlantic Oceans, and ANN3 shows better agreement with the observation than that of ANN1 in some subregions. Significance Statement Global coupled climate models are the primary tool for climate simulation and prediction and provide initial and boundary conditions to drive regional climate models. SST is an essential climate variable simulated and forecast by global climate models, which suffers substantial biases both spatially and temporally. We apply the ensemble averaging of both single- and three-hidden-layer neural networks on the NCEP CFSv2 SST forecast. They can correct the identified SST error, though ANN3 performs relatively better than that of ANN1. Thus, ensemble-based ANN3 is a useful SST bias correction approach.

Evaluating Atmospheric Surface Layer Flux Parameterization within the Coastal Regime

Abstract Traditional atmospheric surface layer theory assumes homogeneous surface conditions. Regardless, nearly all surface layer parameterization schemes employed within numerical weather prediction models utilize the same techniques within highly heterogeneous coastal regimes as for homogeneous environments. We compare predicted surface weather and fluxes of momentum, heat, and moisture—focusing mainly on momentum—from regional simulations using the Coupled Ocean–Atmosphere Mesoscale Prediction System (COAMPS) atmospheric model to observations collected from offshore buoys, inland flux towers, and radiosonde profiles during the Coastal Land-Air-Sea Interaction (CLASI) project throughout the summer of 2021 around Monterey Bay, California. Results reveal that modeled cross-coastal surface flux gradients are spuriously discontinuous, leading to systematically overestimated fluxes and weak winds inland of the coastline during onshore flow periods. Additionally, contrary to observations, modeled surface exchange coefficients are insensitive to wind direction on both sides of the coast, which degrades predictive skill downstream from the coastline. Over the central bay, prediction degrades when near-surface wind directions deviate from the prevailing flow direction as the parameterized stress–wind relationship fails during these cases. Predictive skill over the bay is therefore linked to variations in wind direction. Offshore of the geographically complex peninsula, systematic biases are less clear; however, bifurcations in drag coefficients based on wind direction were measured here as well. Last, increasing the horizontal grid spacing from 333 m to 3 km does not significantly affect surface layer prediction. This work highlights the need to reevaluate surface layer parameterization methods for modeling within coastal regions. Significance Statement Understanding surface layer weather is critical for many purposes, such as infrastructure design and weather forecasting. Within the context of numerical modeling and weather prediction, skillful forecasts of surface winds and temperature rely on accurate portrayal of the surface layer. By comparing observations collected during the Coastal Land-Air-Sea Interaction field program to numerical model solutions, we show that prediction of the surface layer fluxes of momentum, heat, and moisture break down near the coastline, which leads to biases in the predicted surface layer weather both inland and over the water. As surface layer parameterization methods across nearly all numerical models are rooted in the same practices, our results call into question the use of traditional methods near the coastline.

Developing a subseasonal ecological forecast to reduce fisheries bycatch in the Northeast U.S.

Over the past decade, substantial progress has been made in projecting and predicting the spatial distribution of many marine species at seasonal to multidecadal time scales. However, managers and fishers often need to make decisions at much shorter time scales. Subseasonal environmental forecasts, which generate predictions over one to several weeks, can now be combined with species-specific habitat preference data to create ecological forecasts that could facilitate dynamic spatial management. The development of such predictive tools could aid in identifying optimal times and areas for fishers to maximize target catch and avoid nontarget catch. Nontarget catch, or bycatch, can have numerous and potentially severe economic and ecological consequences. Here, we focus on a population of anadromous fish known collectively as river herring (alewife and blueback herring), as they are species of concern and are heavily impacted by bycatch. Using bottom trawl survey data from the Northeast US and subseasonal forecasts of sea surface temperature, we constructed a bycatch risk model to generate probabilistic predictions of river herring distributions in regions frequented by the US mid-water trawl fishery. Assessments of model skill showed that our ecological model performed well in predicting the distribution of river herring and that subseasonal forecasts were effective at 1-week timeframes. There was a clear seasonal effect on forecasted bycatch risk throughout the Northeast US, with particularly high risk in winter and spring months. Importantly, variability in risk was detectable at the weekly timescale and our model identified specific areas and times that fishers should avoid in order to decrease their likelihood of bycatch. The bycatch risk forecast developed in this study is a significant advance from near-real time forecasts and the foundation to build forecast systems by combining species co-occurrence models with subseasonal forecasts. As these subseasonal forecasts are available globally, this approach could be adapted to facilitate the management of other natural resource conflicts around the world.

Improving the accuracy of short-term numerical weather forecasts for the territory of Belarus using the mesoscale WRF model and earth remote sensing data

The problem of improving the WRF numerical weather model performance for the territory of Belarus by assimilating the Earth remote sensing data is considered. It is shown that for the winter period, the use of satellite data of high spatial resolution, including on the structure of land use , albedo, leaf index and photosynthetically active radiation absorbed by the underlying surface can reduce a root-mean-square error of the short-term forecast (up to 48 h) of the air surface temperature by 0.53–1.11 °С. For the summer period, on the basis of numerical experiments the optimal correction factor for the land surface albedo was estimated. This made it possible to reduce a root-mean-square error of temperature forecast at the meteorological stations of Belarus for the lead time of +12, +24, +36, and +48 h by an average of 0.30 °С, 0.10 °С, 0.15 °С, and 0.16 °С, respectively.

Sky View Factor and Screening Impacts on the Forecast Accuracy of Road Surface Temperatures in Finland

Abstract Forecasting road conditions is important, especially in areas with wintry conditions and rapidly changing weather. Accurate forecasts help authorities keep roads safe and optimize maintenance. Considering local features is important when making the forecast because the road surface temperature can vary significantly depending on the road surroundings. For example, in a shadowed location, the road surface temperature can be significantly lower than in open surroundings. A road weather model developed at the Finnish Meteorological Institute is used to forecast the road surface temperature and road conditions. However, the model still assumes open road surroundings. In this study, sky view factor and screening are included in the model, and their effects on the forecast road surface temperature is tested. Road surface temperature hindcasts were performed for 23 selected road weather stations in Finland for three winter periods (October–March) between 2018 and 2021. The results were location dependent, and even changing the lane had a great effect on the verification results in some cases. At best, the screening considerably decreased RMSE values during the day. However, there were many cases in which the screening increased RMSE. In general, the used shadowing algorithm increased the already negative bias during the day. Nevertheless, there were also cases in which the shadowing algorithm improved the bias, especially in February. During the night, the sky view factor made the forecast generally a little warmer, which often slightly decreased the negative bias in the forecast. Significance Statement The screening caused by objects surrounding a road has a great effect on the road surface temperature. Recently, a screening algorithm was added to the Finnish Meteorological Institute’s model that forecasts road conditions. The purpose of this study was to test how the algorithm affects the accuracy of road surface temperature forecasts. According to the results, the screening greatly improved the forecast accuracy in some cases. However, in some cases, the screening made the already overly cold forecast even colder. The study has increased our understanding of the effect of shadowing in the modeled road surface temperatures and helps to create more accurate forecasts in the future.

Postprocessing of Ensemble Weather Forecast Using Decision Tree–Based Probabilistic Forecasting Methods

Abstract Producing an accurate and calibrated probabilistic forecast has high social and economic value. Systematic errors or biases in the ensemble weather forecast can be corrected by postprocessing models whose development is an urgent challenge. Traditionally, the bias correction is done by employing linear regression models that estimate the conditional probability distribution of the forecast. Although this model framework works well, it is restricted to a prespecified model form that often relies on a limited set of predictors only. Most machine learning (ML) methods can tackle these problems with a point prediction, but only a few of them can be applied effectively in a probabilistic manner. The tree-based ML techniques, namely, natural gradient boosting (NGB), quantile random forests (QRF), and distributional regression forests (DRF), are used to adjust hourly 2-m temperature ensemble prediction at lead times of 1–10 days. The ensemble model output statistics (EMOS) and its boosting version are used as benchmark models. The model forecast is based on the European Centre for Medium-Range Weather Forecasts (ECMWF) for the Czech Republic domain. Two training periods 2015–18 and 2018 only were used to learn the models, and their prediction skill was evaluated in 2019. The results show that the QRF and NGB methods provide the best performance for 1–2-day forecasts, while the EMOS method outperforms other methods for 8–10-day forecasts. Key components to improving short-term forecasting are additional atmospheric/surface state predictors and the 4-yr training sample size. Significance Statement Machine learning methods have great potential and are beginning to be widely applied in meteorology in recent years. A new technique called natural gradient boosting (NGB) has been released and used in this paper to refine the probabilistic forecast of surface temperature. It was found that the NGB has better prediction skills than the traditional ensemble model output statistics in forecasting 1 and 2 days in advance. The NGB has similar prediction skills with lower computational demands compared to other advanced machine learning methods such as the quantile random forests. We showed a path to employ the NGB method in this task, which can be followed for refining other and more challenging meteorological variables such as wind speed or precipitation.

Probabilistic extreme SST and marine heatwave forecasts in Chesapeake Bay: A forecast model, skill assessment, and potential value

We test whether skillful 35-day probabilistic forecasts of estuarine sea surface temperature (SST) are possible and whether these forecasts could potentially be used to reduce the economic damages associated with extreme SST events. Using an ensemble of 35-day retrospective forecasts of atmospheric temperature and a simple model that predicts daily mean SST from past SST and forecast atmospheric temperature, we create an equivalent ensemble of retrospective SST forecasts. We compare these SST forecasts with reference forecasts of climatology and damped persistence and find that the SST forecasts are skillful for up to two weeks in the summer. Then, we post-process the forecasts using nonhomogeneous Gaussian regression and assess whether the resulting calibrated probabilistic forecasts are more accurate than the probability implied by the raw model ensemble. Finally, we use an idealized framework to assess whether these probabilistic forecasts can valuably inform decisions to take protective action to mitigate the effects of extreme temperatures and heatwaves. We find that the probabilistic forecasts provide value relative to a naive climatological forecast for 1-2 weeks of lead time, and the value is particularly high in cases where the cost of protection is small relative to the preventable losses suffered when a heatwave occurs. In most cases, the calibrated probabilistic forecasts are also more valuable than deterministic forecasts based on the ensemble mean and naive probabilistic forecasts based on damped persistence. Probabilistic SST forecasts could provide substantial value if applied to adaptively manage the rapid impacts of extreme SSTs, including managing the risks of catch-and-release mortality in fish and Vibrio bacteria in oysters.

An Investigation Into the Variant‐Weight Multi‐Model Ensemble Forecasting Technique Based on the Analyses of Model Systematic Errors Calculation and Elimination

AbstractEliminating the systematic errors of member models is a key step before the variant‐weight multi‐model ensemble forecasting. However, how to reasonably calculate model systematic errors is a problem worthy study. Taking surface temperature as the subject investigated, this study explores this problem through comparative analyses of a series of variant‐weight multi‐model ensemble forecasting experiments. The results showed that eliminating the systematic errors of member models dramatically improves the ensemble forecasting outcomes. The calculation of model systematic errors based on the optimal time lengths regarding the spatiotemporal distribution characteristics of model forecasting errors is a reasonable and effective approach. The surface temperature forecasts with model systematic errors calculated with the optimal time lengths and those beyond the optimal time lengths exhibit higher accuracies. Owing to the approximately uniform distribution of temperature, no significant difference happens between the accuracies of the surface temperature forecasts with model systematic errors calculated using the average and quantile methods. In addition, the updating of the models participating into the ensemble forecasting would weaken the representativeness of the optimal time lengths for model systematic errors calculation that are obtained based on historical period data, and thus, weaken the forecasting skill. The research results provide some train of thought for enhancing the variant‐weight multi‐model ensemble forecasting accuracy.

Prototype of a Warn-on-Forecast System for Smoke (WoFS-Smoke)

Abstract This research begins the process of creating an ensemble-based forecast system for smoke aerosols generated from wildfires using a modified version of the National Severe Storms Laboratory (NSSL) Warn-on-Forecast System (WoFS). The existing WoFS has proven effective in generating short-term (0–3 h) probabilistic forecasts of high-impact weather events such as storm rotation, hail, severe winds, and heavy rainfall. However, it does not include any information on large smoke plumes generated from wildfires that impact air quality and the surrounding environment. The prototype WoFS-Smoke system is based on the deterministic High-Resolution Rapid Refresh-Smoke (HRRR-Smoke) model. HRRR-Smoke runs over a continental United States (CONUS) domain with a 3-km horizontal grid spacing, with hourly forecasts out to 48 h. The smoke plume injection algorithm in HRRR-Smoke is integrated into the WoFS forming WOFS-Smoke so that the advantages of the rapidly cycling, ensemble-based WoFS can be used to generate short-term (0–3 h) probabilistic forecasts of smoke. WoFS-Smoke forecasts from three wildfire cases during 2020 show that the system generates a realistic representation of wildfire smoke when compared against satellite observations. Comparison of smoke forecasts with radar data show that forecast smoke reaches higher levels than radar-detected debris, but exceptions to this are noted. The radiative effect of smoke on surface temperature forecasts is evident, which reduces forecast errors compared to experiments that do not include smoke.

How Skillful Are the European Subseasonal Predictions of Wind Speed and Surface Temperature?

Abstract Subseasonal forecasts of 100-m wind speed and surface temperature, if skillful, can be beneficial to the energy sector as they can be used to plan asset availability and maintenance, assess risks of extreme events, and optimally trade power on the markets. In this study, we evaluate the skill of the European Centre for Medium-Range Weather Forecasts’ subseasonal predictions of 100-m wind speed and 2-m temperature. To the authors’ knowledge, this assessment is the first for the 100-m wind speed, which is an essential variable of practical importance to the energy sector. The assessment is carried out on both forecasts and reforecasts over European domain gridpoint wise and also by considering several spatially averaged domains, using several metrics to assess different attributes of forecast quality. We propose a novel way of synthesizing the continuous ranked probability skill score. The results show that the skill of the forecasts and reforecasts depends on the choice of the climate variable, the period of the year, and the geographical domain. Indeed, the predictions of temperature are better than those of wind speed, with enhanced skill found for both variables in the winter relative to other seasons. The results also indicate significant differences between the skill of forecasts and reforecasts, arising mainly due to the differing ensemble sizes. Overall, depending on the choice of the geographical domain and the forecast attribute, the results show skillful predictions beyond 2 weeks, and in certain cases, up to 6 weeks for both variables, thereby encouraging their implementation in operational decision-making.

Evaluation of WRF-Chem model (v3.9.1.1) real-time air quality forecasts over the Eastern Mediterranean

Abstract. We describe and evaluate a high-resolution real-time air quality forecast system over the Eastern Mediterranean, based on a regional, online coupled atmospheric chemistry and aerosol model. The Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) is used to perform daily, 3 d forecasts of regulated pollutants (NO2, O3, PM2.5) over the Eastern Mediterranean, applying three nested domains with horizontal resolutions of 50, 10 and 2 km, the latter focusing on Cyprus. Natural (dust, sea-salt, biogenic) emissions are calculated online, while anthropogenic emissions are based on the Emissions Database for Global Atmospheric Research – Hemispheric Transport of Air Pollution (EDGAR-HTAP) global emission inventory. A high spatial (1 km) and temporal (hourly) anthropogenic emission inventory is used for the island of Cyprus in the innermost domain. The model skill in forecasting the concentrations of atmospheric pollutants is evaluated using measurements from a network of nine ground stations in Cyprus and compared with the forecasting skill of the EU Copernicus Atmosphere Monitoring Service (CAMS). The forecast of surface temperature, pressure, and wind speed is found to be accurate, with minor discrepancies between the modelled and observed 10 m wind speed at mountainous and coastal sites attributed to the limited representation of the complex topography of Cyprus. Compared to CAMS, the WRF-Chem model predicts with higher accuracy the NO2 mixing ratios at the residential site with a normalized mean bias (NMB) of 7 % during winter and −44 % during summer, whereas the corresponding biases for CAMS are −81 % and −84 %. Due to the high temporal resolution of the anthropogenic emission inventory, the WRF-Chem model captures more accurately the diurnal profiles of NO2 and O3 mixing ratios at the residential site. Background PM2.5 concentrations influenced by long-range transport are overestimated by the WRF-Chem model during winter (NMB = 54 %), whereas the corresponding NMB for CAMS is 11 %. Our results support the adoption of regional, online coupled air quality models over chemical transport models for real-time air quality forecasts.

Исследование многомерного метода еженедельного прогнозирования температуры поверхности Желтого моря на основе кодера-декодера LSTM

In order to further improve the accuracy and stability of sea surface temperature forecasting, this paper uses the 25-year historical data of OISST V2.0 and OAFlux, and fully considers factors such as radiation flux, heat flux, wind speed, air temperature, air specific humidity and SST. By controlling the variables and selecting the best model parameters, a multivariate Yellow Sea SST weekly prediction model based on the Encoder-Decoder LSTM (Long Short Term Memory) was constructed for the first time. The model can effectively track the daily change trend of SST, and respond to its fluctuation changes to achieve relatively accurate prediction. Taking 2008 as an example, the daily absolute errors of the test set within a week are 0.3836, 0.4523, 0.5276, 0.5905, 0.6362, 0.6644, and 0.6827, and the overall RMSE is 0.7594. It is concluded that further research is needed on the optimization of predictors and the ap-plicability of single-point forecasting using the discussed model.

Оценка изменения линейных приростов сосны обыкновенной Печоро-Илычского заповедника на основе климатического прогноза температуры приземного воздуха в российской Арк­тике

For the Arctic region of the European territory of Russia, the authors validated the global climate models of the CMIP6 project with the ERA5 reanalysis data. They identified the most accurate model, based on which made the forecast of surface temperature for the months of July-September under the conditions of two scenarios: high SSP8.5 emissions and low SSP2.6. The results give an idea of the average warming in the next 29 years for these months by 1.8—2.0 (°C). The researchers investigated the response of the increments of the Scots pine (Pinus sylvestris L.) in the Pechora-Ilychsky Reserve to the predicted warming. In addition, they revealed a significant effect of precipitation amounts on the growth parameters of forest stands. The built regression model reflects the total contribution of these climatic factors to the formation of pine increments.

Sea Surface Temperature Forecasting With Ensemble of Stacked Deep Neural Networks

Oceanic temperature has a great impact on global climate and worldwide ecosystems, as its anomalies have been shown to have a direct impact on atmospheric anomalies. The major parameter for measuring the thermal energy of oceans is the sea surface temperature (SST). SST prediction plays an essential role in climatology and ocean-related studies. However, SST prediction is challenging due to the involvement of complex and nonlinear sea thermodynamic factors. To address this challenge, we design a novel ensemble of two stacked deep neural networks (DNNs) that uses air temperature, in addition to water temperature, to improve the SST prediction accuracy. To train our model and compare its accuracy with the state-of-the-art, we employ two well-known datasets from the national oceanic and atmospheric administration as well as the international Argo project. Using DNNs, our proposed method is capable of automatically extracting required features from the input time series and utilizing them internally to provide a highly accurate SST prediction that outperforms state-of-the-art models.

Probabilistic 2-Meter Surface Temperature Forecasting Over Xinjiang Based on Bayesian Model Averaging

Probabilistic 2-Meter Surface Temperature Forecasting Over Xinjiang Based on Bayesian Model Averaging

Hourly road pavement surface temperature forecasting using deep learning models

Road authorities in cold climates regularly apply salt on roads, during winter, to ensure public safety. Pavement surface temperature is a significant parameter affecting snow and ice melting at the onset of a storm. Road temperature below the freezing point of the applied brine causes ice to form on the road surface. Excessive application of salt on the road can have adverse environmental impacts, especially to soil and water quality. Therefore, forecasting pavement temperature can optimize road salt application, while reducing costs, improving public safety, and reducing environmental impacts. This research aims to develop a reliable and accurate pavement surface temperature prediction tool using machine learning techniques. This study employs advanced deep neural network (DNN) learning techniques to predict pavement surface temperature for road salt management purposes. To validate the proposed methodology, this work used hourly solar radiation and air temperature data from Environment Canada, and pavement surface temperature data collected from the Road Weather Information System (RWIS), for sites around the city of Toronto, Ontario. The performance of the proposed DNN model, that integrates a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM), on pavement surface temperature forecasting, was evaluated against four other comparative machine learning methods, LSTM, Convolutional-LSTM (ConvLSTM), Sequence-to-Sequence (Seq2Seq), and Wavelet neural network (Wavenet) models. A dataset of 10,895 samples was collected as an hourly pavement surface temperature record in three timeframes from November 2009 to March 2014 on Highway 401 in southern Ontario, Canada. Experiments included predictions for 1-, 2-, 4- and 6-hours ahead, using as input features air temperature, solar radiation, and present pavement temperature. The proposed pavement temperature forecasting approach resulting from this investigation suggests our new approach is more accurate than previous models. These results reveal that the CNN-LSTM outperforms the four other models and creates predictions closer to the true pavement surface temperature.