Reconstructing and Forecasting Environmental Data with Neural Networks
In this project, I addressed a common limitation in environmental monitoring: the sparse temporal overlap between in‑situ observations (suspended sediment concentrations) and satellite remote‑sensing data. To overcome this gap, I implemented a sequential neural network using the TensorFlow Keras library, leveraging multi‑parametric inputs — including meteorological conditions, wave dynamics, and riverine discharge — to hindcast and reconstruct missing observations across time.

This approach not only improved the continuity of long‑term environmental datasets but also laid the foundation for forecasting future dynamics. Compared to traditional multivariate linear regression, neural networks excel in capturing hidden, nonlinear relationships within complex environmental data. By learning from patterns and interactions that would otherwise be overlooked, this method provides a more accurate and robust representation of sediment dynamics across a range of temporal and environmental conditions. To optimize performance, I tuned key hyperparameters to balance model complexity, prediction accuracy, and generalization.
I also explored the use of Long Short‑Term Memory (LSTM) architectures to explicitly encode temporal dependencies and long‑term persistence within the data. However, due to the limited availability of prior information and short record length, I opted for a more general sequential approach for this application. Nevertheless, this work highlights the potential of deep learning to fill temporal gaps in environmental observations and enable forward‑looking forecasting — making it a valuable tool for understanding and managing coastal dynamics in a changing climate.