Design of an Iterative Model Integrating Temporal Convolutional Networks and Transfer Learning for Hydrological Analysis

Advanced artificial intelligence with hydrological analysis promotes accurate Water-Level (WL) forecasting and anomaly detection, which are crucial for disaster management, especially in frequent flood-prone Kerala. This research presents a Seasonal Hybrid Extreme Studentized Deviation Test (SHESD), and transfer learning based Long Short-Term Memory (LSTM) with advanced technique Temporal Convolutional Networks (TCN) in modelling temporal patterns. By incorporating the SHESD, seasonal anomalies will be effectively found. The residuals were measured using ESD by classifying timestamp series into seasonal and trend dataset NDWI downloaded from GEE. The method detects abnormal WL occurrences that indicate significant flood consequences for diverse scenarios with 90% high accuracy. WL predictions are accurately measured using the proposed model with 87% R2, 0.095 mean absolute error and 0.03 Mean Squared Error. Fine-tuning these integrated deep learning models using local data enhanced the model prediction accuracy.