Predicting spatio-temporal concentrations of PM<Subscript>2.5</Subscript> using land use and meteorological data in Yangtze River Delta,China

The prediction of PM_2.5 concentrations with high spatiotemporal resolution has been suggested as a potential method for data collection to assess the health effects of exposure. This work predicted the weekly average PM_2.5 concentrations in the Yangtze River Delta, China, by using a spatio-temporal model. Integrating land use data, including the areas of cultivated land, construction land, and forest land, and meteorological data, including precipitation, air pressure, relative humidity, temperature, and wind speed, we used the model to estimate the weekly average PM_2.5 concentrations. We validated the estimated effects by using the cross-validated R² and Root mean square error (RMSE); the results showed that the model performed well in capturing the spatiotemporal variability of PM_2.5 concentration, with a reasonably large R² of 0.86 and a small RMSE of 8.15 (μg/m³). In addition, the predicted values covered 94% of the observed data at the 95% confidence interval. This work provided a dataset of PM_2.5 concentration predictions with a spatiotemporal resolution of 3 km × week, which would contribute to accurately assessing the potential health effects of air pollution.     

Forecasting Hourly Roadside Particulate Matter in Taipei County of Taiwan Based on First‐Order and One‐Variable Grey Model

In this study, seven types of first‐order and one‐variable grey differential equation model (abbreviated as GM (1, 1) model) were used to forecast hourly roadside particulate matter (PM) including PM₁₀ and PM_2.5 concentrations in Taipei County of Taiwan. Their forecasting performance was also compared. The results indicated that the minimum mean absolute percentage error (MAPE), mean squared error (MSE), root mean squared error (RMSE), and maximum correlation coefficient (R) was 11.70%, 60.06, 7.75, and 0.90%, respectively when forecasting PM₁₀. When forecasting PM_2.5, the minimum MAPE, MSE, RMSE, and maximum R‐value of 16.33%, 29.78, 5.46, and 0.90, respectively could be achieved. All statistical values revealed that the forecasting performance of GM (1, 1, x⁽⁰⁾), GM (1, 1, a), and GM (1, 1, b) outperformed other GM (1, 1) models. According to the results, it revealed that GM (1, 1) was an efficiently early warning tool for providing PM information to the roadside inhabitants.     

Estimation of PM<Subscript>10</Subscript> concentration from air quality data in the vicinity of a major steelworks site in the metropolitan area of Avilés (Northern Spain) using machine learning techniques

Atmospheric particulate matter (PM) is one of the pollutants that may have a significant impact on human health. Data collected over 7 years from the air quality monitoring station at the LD-III steelworks, belonging to the Arcelor-Mittal Steel Company, located in the metropolitan area of Avilés (Principality of Asturias, Northern Spain), is analyzed using four different mathematical models: vector autoregressive moving-average, autoregressive integrated moving-average (ARIMA), multilayer perceptron neural networks and support vector machines with regression. Measured monthly, the average concentration of pollutants (SO₂, NO and NO₂) and PM₁₀ (particles with a diameter less than ?10 μm) is used as input to forecast the monthly average concentration of PM₁₀ from one to 7 months ahead. Simulations showed that the ARIMA model performs better than the other models when forecasting 1 month ahead, while in the forecast from one to 9 months ahead the best performance is given by the support vector regression.     

Integrating hydrometeorological information for rainfall‐runoff modelling by artificial neural networks

The major purpose of this study is to effectively construct artificial neural networks‐based multistep ahead flood forecasting by using hydrometeorological and numerical weather prediction (NWP) information. To achieve this goal, we first compare three mean areal precipitation forecasts: radar/NWP multisource‐derived forecasts (P_r), NWP precipitation forecasts (P_n), and improved precipitation forecasts (P_m) by merging P_r and P_n. The analysis shows that the accuracy of P_m is higher than that of P_r and P_n. The analysis also indicates that the NWP precipitation forecasts do provide relative effectiveness to the merging procedure, particularly for forecast lead time of 4–6 h. In sum, the merged products performed well and captured the main tendency of rainfall pattern. Subsequently, a recurrent neural network (RNN)‐based multistep ahead flood forecasting techniques is produced by feeding in the merged precipitation. The evaluation of 1–6‐h flood forecasting schemes strongly shows that the proposed hydrological model provides accurate and stable flood forecasts in comparison with a conventional case, and significantly improves the peak flow forecasts and the time‐lag problem. An important finding is the hydrologic model responses which do not seem to be sensitive to precipitation predictions in lead times of 1–3 h, whereas the runoff forecasts are highly dependent on predicted precipitation information for longer lead times (4–6 h). Overall, the results demonstrate that accurate and consistent multistep ahead flood forecasting can be obtained by integrating predicted precipitation information into ANNs modelling. Copyright © 2009 John Wiley & Sons, Ltd.     

The predictability of near-coastal currents using a baroclinic unstructured grid model

A limited domain, coastal ocean forecast system consisting of an unstructured grid model, a meteorological model, a regional ocean model, and a global tidal database is designed to be globally relocatable. For such a system to be viable, the predictability of coastal currents must be well understood with error sources clearly identified. To this end, the coastal forecast system is applied at the mouth of Chesapeake Bay in response to a Navy exercise. Two-day forecasts are produced for a 10-day period from 4 to 14 June 2010 and compared to real-time observations. Interplay between the temporal frequency of the regional model boundary forcing and the application of external tides to the coastal model impacts the tidal characteristics of the coastal current, even contributing a small phase error. Frequencies of at least 3 h are needed to resolve the tidal signal within the regional model; otherwise, externally applied tides from a database are needed to capture the tidal variability. Spatial resolution of the regional model (3 vs 1 km) does not impact skill of the current prediction. Tidal response of the system indicates excellent representation of the dominant M ₂ tide for water level and currents. Diurnal tides, especially K ₁, are amplified unrealistically with the application of coarse 27-km winds. Higher-resolution winds reduce current forecast error with the exception of wind originating from the SSW, SSE, and E. These winds run shore parallel and are subject to strong interaction with the shoreline that is poorly represented even by the 3-km wind fields. The vertical distribution of currents is also well predicted by the coastal model. Spatial and temporal resolution of the wind forcing including areas close to the shoreline is the most critical component for accurate current forecasts. Additionally, it is demonstrated that wind resolution plays a large role in establishing realistic thermal and density structures in upwelling prone regions.     

Short‐term inflow forecasting using an artificial neural network model

The primary objective of this study is to investigate the possibility of including more temporal and spatial information on short‐term inflow forecasting, which is not easily attained in the traditional time‐series models or conceptual hydrological models. In order to achieve this objective, an artificial neural network (ANN) model for short‐term inflow forecasting is developed and several issues associated with the use of an ANN model are examined in this study. The formulated ANN model is used to forecast 1‐ to 7‐h ahead inflows into a hydropower reservoir. The root‐mean‐squared error (RMSE), the Nash–Sutcliffe coefficient (NSC), the A information criterion (AIC), B information criterion (BIC) of the 1‐ to 7‐h ahead forecasts, and the cross‐correlation coefficient between the forecast and observed inflows are estimated. Model performance is analysed and some quantitative analysis is presented. The results obtained are satisfactory. Perceived strengths of the ANN model are the capability for representing complex and non‐linear relationships as well as being able to include more information in the model easily. Although the results obtained may not be universal, they are expected to reveal some possible problems in ANN models and provide some helpful insights in the development and application of ANN models in the field of hydrology and water resources. Copyright © 2002 John Wiley & Sons, Ltd.     

Comparing quantitative precipitation forecast methods for prediction of sewer flows in a small urban area

Abstract

Due to the relatively small spatial scale, as well as rapid response, of urban drainage systems, the use of quantitative rainfall forecasts for providing quantitative flow and depth predictions is a challenging task. Such predictions are important when consideration is given to urban pluvial flooding and receiving water quality, and it is worthwhile to investigate the potential for improved forecasting. In this study, three quantitative precipitation forecast methods of increasing complexity were compared and used to create quantitative forecasts of sewer flows 0–3 h ahead in the centre of a small town in the north of England. The HyRaTrac radar nowcast model was employed, as well as two different versions of the more complex STEPS model. The STEPS model was used as a deterministic nowcasting system, and was also blended with the Numerical Weather Prediction (NWP) model MM5 to investigate the potential of increasing forecast lead-times (LTs) using high-resolution NWP. Predictive LTs between 15 and 90 min gave acceptable results, but were a function of the event type. It was concluded that higher resolution rainfall estimation as well as nowcasts are needed for prediction of both local pluvial flooding and combined sewer overflow spill events.

Editor D. Koutsoyiannis; Guest editor R.J. Moore     

The December 26, 2004, off the west coast of northern Sumatra,Indonesia, <Emphasis Type="Italic">M</Emphasis><Subscript>W</Subscript>=9.0, earthquake and the critical-point-like model of earthquake preparation

Long-term seismic activity prior to the December 26, 2004, off the west coast of northern Sumatra, Indonesia, M _W=9.0 earthquake was investigated using the Harvard CMT catalogue. It is observed that before this great earthquake, there exists an accelerating moment release (AMR) process with the temporal scale of a quarter century and the spatial scale of 1 500 km. Within this spatial range, the M _W=9.0 event falls into the piece-wise power-law-like frequency-magnitude distribution. Therefore, in the perspective of the critical-point-like model of earthquake preparation, the failure to forecast/predict the approaching and/or the size of this earthquake is not due to the physically intrinsic unpredictability of earthquakes.     

Short term streamflow forecasting using artificial neural networks

The research described in this article investigates the utility of Artificial Neural Networks (ANNs) for short term forecasting of streamflow. The work explores the capabilities of ANNs and compares the performance of this tool to conventional approaches used to forecast streamflow. Several issues associated with the use of an ANN are examined including the type of input data and the number, and the size of hidden layer(s) to be included in the network. Perceived strengths of ANNs are the capability for representing complex, non-linear relationships as well as being able to model interaction effects. The application of the ANN approach is to a portion of the Winnipeg River system in Northwest Ontario, Canada. Forecasting was conducted on a catchment area of approximately 20 000 km². using quarter monthly time intervals. The results were most promising. A very close fit was obtained during the calibration (training) phase and the ANNs developed consistently outperformed a conventional model during the verification (testing) phase for all of the four forecast lead-times. The average improvement in the root mean squared error (RMSE) for the 8 years of test data varied from 5 cms in the four time step ahead forecasts to 12.1 cms in the two time step ahead forecasts.     

Compositional time series analysis for Air Pollution Index data

The increasing importance of understanding the structure of Air Pollution Index (API) makes it necessary to come out with a compositional of API based on its pollutants. This will be more comprehensible for the public and easier to cooperate with authorities in reducing the causes of air pollution. Since five pollutants contribute in determining the API values, API can be shown as a compositional data. This study is conducted based on the data of API and its pollutants collected from Klang city in Malaysia for the period of January 2005 to December 2014. The proportion of each pollutant in API is considered as a component with five components in a compositional API. The existence of zero components in some pollutants, that have no effect on API, is a serious problem that prevents the application of log-ratio transformation. Thus, the approach of amalgamation has been used to combine the components with zero in order to reduce the number of zeros. Also, a multiplicative replacement has been utilized to eliminate the zero components and replace them with a small value that maintains the ratios of nonzero components. Transforming the compositional data to log-ratio coordinates has been done using the additive log ratio transformation, and the transformed series is then modeled by using a VAR model. Four criteria are used to determine the number of lags p of VAR(p) and these are: the Akaike Information, the Schwartz, the Hannan–Quinn and the Final Prediction Error criteria. Based on the results, A VAR (1) model with no constants or trend is considered as the best fitted model and it is used to forecast 12 months ahead. In addition, API values are mainly determined by PM₁₀ that has a proportion close to one most of the time during study period. Therefore, authorities and researchers need to study the sources of PM₁₀ and provide the public with useful information and alternatives in term of reducing the air pollution.     

Three-dimensional analysis of ozone and PM<Subscript>2.5</Subscript> distributions obtained by observations of tethered balloon and unmanned aerial vehicle in Shanghai,China

A lightweight unmanned aerial vehicle (UAV) and a tethered balloon platform were jointly used to investigate three-dimensional distributions of ozone and PM_2.5 concentrations within the lower troposphere (1000 m) at a localized coastal area in Shanghai, China. Eight tethered balloon soundings and three UAV flights were conducted on May 25, 2016. Generalized additive models (GAMs) were used to quantitatively describe the relationships between air pollutants and other obtained parameters. Field observations showed that large variations were captured both in the vertical and horizontal distributions of ozone and PM_2.5 concentrations. Significant stratified layers of ozone and PM_2.5 concentrations as well as wind directions were observed throughout the day. Estimated bulk Richardson numbers indicate that the vertical mixing of air masses within the lower troposphere were heavily suppressed throughout the day, leading to much higher concentrations of ozone and PM_2.5 in the planetary boundary layer (PBL). The NO and NO₂ concentrations in the experimental field were much lower than that in the urban area of Shanghai and demonstrated totally different vertical distribution patterns from that of ozone and PM_2.5. This indicates that aged air masses of different sources were transported to the experimental field at different heights. Results derived from the GAMs showed that the aggregate impact of the selected variables for the vertical variations can explain 94.3% of the variance in ozone and 94.5% in PM_2.5. Air temperature, relative humidity and atmospheric pressure had the strongest effects on the variations of ozone and PM_2.5. As for the horizontal variations, the GAMs can explain 56.3% of the variance in ozone and 57.6% in PM_2.5. The strongest effect on ozone was related to air temperature, while PM_2.5 was related to relative humidity. The output of GAMs also implied that fine aerosol particles were in the stage of growth in the experimental field, which is different from ozone (aged air parcels of ozone). Geographical parameters influenced the horizontal variations of ozone and PM_2.5 concentrations by changing underlying surface types. The differences of thermodynamic properties between land and sea resulted in quick changes of PBL height, air temperature and dew point over the coastal area, which was linked to the extent of vertical mixing at different locations. The results of GAMs can be used to analyze the sources and formation mechanisms of ozone and PM_2.5 pollutions at a localized area.     

The factors influencing the abrasion efficiency of saltating grains on a clay‐crusted playa

The entrainment and subsequent transport of PM₁₀ (particulate matter <10 µm) has become an important and challenging focus of research for both scientific and practical applications. Arid and semi‐arid environments are important sources for the atmospheric loading of PM₁₀, although the emission of this material is often limited by surface crusts. It has been suggested that the primary mechanisms through which PM₁₀ is released from a crusted surface are abrasion by saltating grains or disturbance by agricultural and recreational activities. To examine the importance of saltation abrasion in the emission of PM₁₀, a series of field wind tunnel tests were conducted on a clay‐crusted surface near Desert Wells, Arizona. In a previous part of this study it was found that the emission rate varies linearly with the saltation transport rate, although there can be considerable variation in this relationship. This paper more closely examines the source of the variability in the abrasion efficiency, the amount of PM₁₀ emitted by a given quantity of saltating grains. The abrasion efficiency was found to vary with the susceptibility of the surface to abrasion, the ability of the sand to abrade that surface and the availability of material with a caliper size <10 µm within the crust. Specifically, the results of the study show that the abrasion efficiency is related to the crust strength, the amount of surface disturbance and the velocity of the saltating grains. It is concluded that the spatial and temporal variability of these controls on the abrasion efficiency imposes severe contextual limitations on experimentally derived models, and can make theoretical models too complex and impractical to be of use. Copyright­© 2001 John Wiley & Sons, Ltd.     

Updating Real-Time Flood Forecasting Using a Fuzzy Rule-Based Model/Mise à Jour de Prévision de Crue en Temps Réel Grâce à un Modèle à Base de Règles Floues

Abstract

An updating technique is a tool to update the forecasts of mathematical flood forecasting model based on data observed in real time, and is an important element in a flood forecasting model. An error prediction model based on a fuzzy rule-based method was proposed as the updating technique in this work to improve one- to four-hour-ahead flood forecasts by a model that is composed of the grey rainfall model, the grey rainfall—runoff model and the modified Muskingum flow routing model. The coefficient of efficiency with respect to a benchmark is applied to test the applicability of the proposed fuzzy rule-based method. The analysis reveals that the fuzzy rule-based method can improve flood forecasts one to four hours ahead. The proposed updating technique can mitigate the problem of the phase lag in forecast hydrographs, and especially in forecast hydrographs with longer lead times.     

Source Parameters of the Deadly M<Subscript>w</Subscript> 7.6 Kashmir Earthquake of 8 October, 2005

During the last six years, National Geophysical Research Institute, Hyderabad has established a semi-permanent seismological network of 5–8 broadband seismographs and 10–20 accelerographs in the Kachchh seismic zone, Gujarat with a prime objective to monitor the continued aftershock activity of the 2001 M_w 7.7 Bhuj mainshock. The reliable and accurate broadband data for the 8 October M_w 7.6 2005 Kashmir earthquake and its aftershocks from this network as well as Hyderabad Geoscope station enabled us to estimate the group velocity dispersion characteristics and one-dimensional regional shear velocity structure of the Peninsular India. Firstly, we measure Rayleigh-and Love-wave group velocity dispersion curves in the period range of 8 to 35 sec and invert these curves to estimate the crustal and upper mantle structure below the western part of Peninsular India. Our best model suggests a two-layered crust: The upper crust is 13.8 km thick with a shear velocity (Vs) of 3.2 km/s; the corresponding values for the lower crust are 24.9 km and 3.7 km/sec. The shear velocity for the upper mantle is found to be 4.65 km/sec. Based on this structure, we perform a moment tensor (MT) inversion of the bandpass (0.05–0.02 Hz) filtered seismograms of the Kashmir earthquake. The best fit is obtained for a source located at a depth of 30 km, with a seismic moment, Mo, of 1.6 × 10²⁷ dyne-cm, and a focal mechanism with strike 19.5°, dip 42°, and rake 167°. The long-period magnitude (M_A ~ M_w) of this earthquake is estimated to be 7.31. An analysis of well-developed sPn and sSn regional crustal phases from the bandpassed (0.02–0.25 Hz) seismograms of this earthquake at four stations in Kachchh suggests a focal depth of 30.8 km.     

River temperature forecasting: case study for Little Southwest Miramichi River (New Brunswick,Canada)

River temperature models play an increasingly important role in the management of fisheries and aquatic resources. Among river temperature models, forecasting models remain relatively unused compared to water temperature simulation models. However, water temperature forecasting is extremely important for in-season management of fisheries, especially when short-term forecasts (a few days) are required. In this study, forecast and simulation models were applied to the Little Southwest Miramichi River (New Brunswick, Canada), where water temperatures can regularly exceed 25–29°C during summer, necessitating associated fisheries closures. Second- and third-order autoregressive models (AR2, AR3) were calibrated and validated using air temperature as the exogenous variable to predict minimum, mean and maximum daily water temperatures. These models were then used to predict river temperatures in forecast mode (1-, 2- and 3-day forecasts using real-time data) and in simulation mode (using only air temperature as input). The results showed that the models performed better when used to forecast rather than simulate water temperatures. The AR3 model slightly outperformed the AR2 in the forecasting mode, with root mean square errors (RMSE) generally between 0.87°C and 1.58°C. However, in the simulation mode, the AR2 slightly outperformed the AR3 model (1.25°C < RMSE < 1.90°C). One-day forecast models performed the best (RMSE ~ 1°C) and model performance decreased as time lag increased (RMSE close to 1.5°C after 3 days). The study showed that marked improvement in the modelling can be accomplished using forecasting models compared to water temperature simulations, especially for short-term forecasts.

EDITOR M.C. Acreman ASSOCIATE EDITOR S. Huang     

The impact of climate change on sea level rise at Peninsular Malaysia and Sabah–Sarawak

The sea level change along the Peninsular Malaysia and Sabah–Sarawak coastlines for the 21st century is investigated along the coastal areas of Peninsular Malaysia and Sabah–Sarawak because of the expected climate change during the 21st century. The spatial variation of the sea level change is estimated by assimilating the global mean sea level projections from the Atmosphere–Ocean coupled Global Climate Model/General Circulation Model (AOGCM) simulations to the satellite altimeter observations along the subject coastlines. Using the assimilated AOGCM projections, the sea level around the Peninsular Malaysia coastline is projected to rise with a mean in the range of 0.066 to 0.141 m in 2040 and 0.253 m to 0.517 m in 2100. Using the assimilated AOGCM projections, the sea level around Sabah–Sarawak coastlines is projected to rise with a mean in the range of 0.115 m to 0.291 m in 2040 and 0.432 m to 1.064 m in 2100. The highest sea level rise occurs at the northeast and northwest regions in Peninsular Malaysia and at north and east sectors of Sabah in Sabah–Sarawak coastline. Copyright © 2012 John Wiley & Sons, Ltd.     

Modelling soil moisture in a high‐latitude landscape using LiDAR and soil data

Soil moisture has a pronounced effect on earth surface processes. Global soil moisture is strongly driven by climate, whereas at finer scales, the role of non‐climatic drivers becomes more important. We provide insights into the significance of soil and land surface properties in landscape‐scale soil moisture variation by utilizing high‐resolution light detection and ranging (LiDAR) data and extensive field investigations. The data consist of 1200 study plots located in a high‐latitude landscape of mountain tundra in north‐western Finland. We measured the plots three times during growing season 2016 with a hand‐held time‐domain reflectometry sensor. To model soil moisture and its temporal variation, we used four statistical modelling methods: generalized linear models, generalized additive models, boosted regression trees, and random forests. The model fit of the soil moisture models were R² = 0.60 and root mean square error (RMSE) 8.04 VWC% on average, while the temporal variation models showed a lower fit of R² = 0.25 and RMSE 13.11 CV%. The predictive performances for the former were R² = 0.47 and RMSE 9.34 VWC%, and for the latter R² = 0.01 and RMSE 15.29 CV%. Results were similar across the modelling methods, demonstrating a consistent pattern. Soil moisture and its temporal variation showed strong heterogeneity over short distances; therefore, soil moisture modelling benefits from high‐resolution predictors, such as LiDAR based variables. In the soil moisture models, the strongest predictor was SAGA (System for Automated Geoscientific Analyses) wetness index (SWI), based on a 1 m² digital terrain model derived from LiDAR data, which outperformed soil predictors. Thus, our study supports the use of LiDAR based SWI in explaining fine‐scale soil moisture variation. In the temporal variation models, the strongest predictor was the field‐quantified organic layer depth variable. Our results show that spatial soil moisture predictions can be based on soil and land surface properties, yet the temporal models require further investigation. Copyright © 2017 John Wiley & Sons, Ltd.     

The data‐driven approach as an operational real‐time flood forecasting model

Accurate water level forecasts are essential for flood warning. This study adopts a data‐driven approach based on the adaptive network–based fuzzy inference system (ANFIS) to forecast the daily water levels of the Lower Mekong River at Pakse, Lao People's Democratic Republic. ANFIS is a hybrid system combining fuzzy inference system and artificial neural networks. Five ANFIS models were developed to provide water level forecasts from 1 to 5 days ahead, respectively. The results show that although ANFIS forecasts of water levels up to three lead days satisfied the benchmark, four‐ and five‐lead‐day forecasts were only slightly better in performance compared with the currently adopted operational model. This limitation is imposed by the auto‐ and cross‐correlations of the water level time series. Output updating procedures based on the autoregressive (AR) and recursive AR (RAR) models were used to enhance ANFIS model outputs. The RAR model performed better than the AR model. In addition, a partial recursive procedure that reduced the number of recursive steps when applying the AR or the RAR model for multi‐step‐ahead error prediction was superior to the fully recursive procedure. The RAR‐based partial recursive updating procedure significantly improved three‐, four‐ and five‐lead‐day forecasts. Our study further shows that for long lead times, ANFIS model errors are dominated by lag time errors. Although the ANFIS model with the RAR‐based partial recursive updating procedure provided the best results, this method was able to reduce the lag time errors significantly for the falling limbs only. Improvements for the rising limbs were modest. Copyright © 2011 John Wiley & Sons, Ltd.     

Mass fraction spatiotemporal geostatistics and its application to map atmospheric polycyclic aromatic hydrocarbons after 9/11

This work proposes a space/time estimation method for atmospheric PM_2.5 components by modelling the mass fraction at a selection of space/time locations where the component is measured and applying the model to the extensive PM_2.5 monitoring network. The method we developed utilizes the nonlinear Bayesian maximum entropy framework to perform the geostatistical estimation. We implemented this approach using data from nine carcinogenic, particle-bound polycyclic aromatic hydrocarbons (PAHs) measured from archived PM_2.5 samples collected at four locations around the World Trade Center (WTC) from September 22, 2001 to March 27, 2002. The mass fraction model developed at these four sites was used to estimate PAH concentrations at additional PM_2.5 monitors. Even with limited PAH data, a spatial validation showed the application of the mass fraction model reduced the mean squared error (MSE) by 7–22%, while in the temporal validation there was an exponential improvement in MSE positively associated with the number of days of PAH data removed. Our results include space/time maps of atmospheric PAH concentrations in the New York area after 9/11.     

Deep learning approaches for improving prediction of daily stream temperature in data-scarce,unmonitored, and dammed basins

Basin-centric long short-term memory (LSTM) network models have recently been shown to be an exceptionally powerful tool for stream temperature (T_s) temporal prediction (training in one period and predicting in another period at the same sites). However, spatial extrapolation is a well-known challenge to modelling T_s and it is uncertain how an LSTM-based daily T_s model will perform in unmonitored or dammed basins. Here we compiled a new benchmark dataset consisting of >400 basins across the contiguous United States in different data availability groups (DAG, meaning the daily sampling frequency) with and without major dams, and studied how to assemble suitable training datasets for predictions in basins with or without temperature monitoring. For prediction in unmonitored basins (PUB), LSTM produced a root-mean-square error (RMSE) of 1.129°C and an R² of 0.983. While these metrics declined from LSTM's temporal prediction performance, they far surpassed traditional models' PUB values, and were competitive with traditional models' temporal prediction on calibrated sites. Even for unmonitored basins with major reservoirs, we obtained a median RMSE of 1.202°C and an R² of 0.984. For temporal prediction, the most suitable training set was the matching DAG that the basin could be grouped into (for example, the 60% DAG was most suitable for a basin with 61% data availability). However, for PUB, a training dataset including all basins with data was consistently preferred. An input-selection ensemble moderately mitigated attribute overfitting. Our results indicate there are influential latent processes not sufficiently described by the inputs (e.g., geology, wetland covers), but temporal fluctuations can still be predicted well, and LSTM appears to be a highly accurate T_s modelling tool even for spatial extrapolation.