Pavitra Kumar’s research while affiliated with University of Liverpool and other places

Understanding and predicting coastal change is of the foremost importance to protect coastal communities and coastal assets. This study analyzes field data from 125 locations along the Morecambe coastline, consisting of beach transects collected twice a year for more than a decade (2007–2022). Wave data at these 125 locations were simulated using the hydrodynamic Delft3D model, with full coupling of the Delft3D FLOW and WAVE modules. To model the sediment volume changes observed along the Morecambe coastline, this study proposes a two‐stage machine learning model that incorporates beach behavior classification and deep learning techniques to predict changes in sediment volumes along coastal environments. The first stage of the model, developed using a random forest classifier, classifies beach behavior into four categories: eroding, accreting, stable, or undergoing short‐term fluctuations. The second stage of the model developed using LSTM and sequence‐to‐sequence models, uses the output of the first stage to predict the change in sediment volume after erosion/accretion. The random forest classifier achieves testing accuracy of 0.74. LSTM model achieved a testing regression of 0.92 for one‐step‐ahead (6 months) predictions of change in sediment volume time series, while sequence‐to‐sequence model achieved the testing regression of 0.96 for three‐time‐ahead (1.5 years) predictions and 0.88 for ten‐time‐step‐ahead (5 years) prediction.

Our coasts are facing a growing threat from rising sea levels and detrimental storm events. Nature-based solutions are gaining interest for their potential to provide multiple ecosystem services, including coastal protection. Seagrass can influence coastal sediment transport and wave propagation, however, whether seagrass can be used as an effective intervention for coastal protection is unclear. Using the Delft3D model, this study looks at how seagrass affects coastal hydrodynamics in a macrotidal bay, under idealised scenarios that simulate a rise in sea level and storm wave heights, emulating projected sea conditions under climate change. Through the use of a habitat suitability model, a seagrass patch in Morecambe Bay was simulated, based on a location within the bay that is most suitable for seagrass species Zostera marina. Hydrodynamic simulations were run for a number of scenarios for a domain with and without a seagrass patch. Scenarios included variable boundary wave heights, sea levels and vegetation parameters, to test the influence of seagrass in future climates. Results show that there is a reduction in mean and maximum wave height inside and behind the seagrass patch, with the largest changes in wave dissipation observed during higher boundary wave conditions. In these conditions the maximum wave height reduced by over a half within the patch. Simulations examining the influence of seagrass density and canopy height found that a lower density seagrass patch reduced maximum wave height more substantially than higher density patches. Although alternative hard engineered solutions are able to attenuate wave energy more effectively than seagrass, these results demonstrate seagrass could play an important role in conjunction with salt marsh restoration or existing hard engineering solutions, helping to mitigate the risk of flooding and erosion, whilst providing additional ecosystem services such as carbon sequestration and habitat provision.

Europe is facing a novel demographic shift in population decline. Across the continent an increasing number of countries and sub-national areas are undergoing trajectories of depopulation, threatening to challenge the functioning of societies. Understanding the likely population futures is essential in preparing for such unprecedented population change, yet obtaining population forecasts at granular geographic scales remains challenging. Population forecasts are often only produced at large spatial scales, owing to the requirement of highly detailed demographic data, which are typically lacking for small areas. However, recent advancements in remote sensing have facilitated geographically granular demographic estimates, enabling the production of high-resolution population forecasts. Further developments in machine-learning based forecasting methodologies present an exciting opportunity to predict demographic futures. Making use of WorldPop gridded population count data, we develop a Long Short-Term Memory (LSTM) recurrent neural network to forecast municipal level population change across the entire European continent. In search of optimal parameters for population forecasting, a series of models are developed that differ by data specification, architecture, and implementation approach. Their evaluation highlights the advantages of models with a singular LSTM layer, longer input sequences, subset implementation and explicit consideration of spatial population dynamics resulting in improved forecast accuracy. Our best performing model, with a median absolute percentage error of 1.29, is further developed to predict future population change outcomes to 2030. The forecast reveals that population decline will be affecting 70.2% of European municipalities by 2030, and a majority of countries (75%), directly implicating over half (59.3%) of the European population.

Within the context of climate change, understanding and predicting coastal change is of the foremost importance to protect coastal communities and coastal assets. This study analyzes field data from 125 locations along the Morecambe coastline, consisting of beach transects collected twice a year for more than a decade (2007 to 2022). To model the sediment volume changes observed along the Morecambe coastline, this study proposes a two-stage machine learning model that incorporates beach behavior classification and deep learning techniques to predict changes in sediment volumes along coastal environments. The first stage of the model, developed using a random forest classifier, classifies beach behavior into four categories: eroding, accreting, stable, or undergoing short-term fluctuations. The second stage of the model developed using LSTM and sequence-to-sequence models, uses the output of the first stage to predict the available sediment volume after erosion/accretion. LSTM model achieved a testing regression of 0.9961 for one-step-ahead (6 months) predictions of sediment volume time series, while sequence-to-sequence model achieved the testing regression of 0.9950 for three-time-ahead (1.5 years) predictions and 0.9916 for ten-time-step-ahead (5 years) prediction.

Coastal protection is of paramount importance because erosion and flooding affect millions of people living along the coast and can largely influence countries' economy. The implementation of nature‐based solutions for coastal protection, such as sand engines, has become more popular due to these interventions' adaptability to climate change. This study explores synergies between Artificial Intelligence (AI) and hydro‐morphodynamic models for the creation of efficient decision‐making tools for the choice of optimal sand engines configurations. Specifically, we investigate the use of long‐short‐term memory (LSTM) models as predictive tools for the morphological evolution of sand engines. We developed different LSTM models to predict time series of bathymetric changes across the sand engine as well as the time‐decline in the sand engine volume as a function of external forces and intervention size. Finally, a MATLAB framework was developed to return LSTM model results based on users' inputs about sand engine size and external forcings.

Accurate prediction of short-term water demand, especially, in the case of extreme weather conditions such as flood, droughts and storms, is crucial information for the policy makers to manage the availability of freshwater. This study develops a hybrid model for the prediction of monthly water demand using the database of monthly urban water consumption in Melbourne, Australia. The dataset consisted of minimum, maximum, and mean temperature (°C), evaporation (mm), rainfall (mm), solar radiation (MJ/m 2), maximum relative humidity (%), vapor pressure (hpa), and potential evapotranspiration (mm). The dataset was normalized using natural logarithm and denoized then by employing the discrete wavelet transform. Principle component analysis was used to determine which predictors were most reliable. Hybrid model development included the optimization of ANN coefficients (its weights and biases) using adaptive guided differential evolution algorithm. Post-optimization ANN model was trained using eleven different leaning algorithms. Models were trained several times with different configuration (nodes in hidden layers) to achieve better accuracy. The final optimum learning algorithm was selected based on the performance values (regression; mean absolute, relative and maximum error) and Taylor diagram.