ViT-A-LSTM-DE Framework With Dense Fusion Encoder for Superior Algal Bloom Risk Assessment and Monitoring

The occurrence of algal blooms poses serious environmental, economic, and health risks, requiring reliable and timely forecasting systems. This research introduces a multimodel framework named modified vision transformer-based adaptive long short term memory with dynamic encoder for accurate detection and prediction of algal bloom. A modified vision transformer is employed to extract the fine-grained spatial features from satellite images while an adaptive long short-term memory with dynamic encoder captures temporal dependencies and contextual variations in time-series data. These spatial and temporal features are integrated using a dense transformer fusion encoder method, which then fuses spatial and temporal representations into a unified cross-modal embedding. Additionally, a deep autoencoder is used to identify anomalies in water quality parameters, with anomaly scores serving as an auxiliary input to improve prediction accuracy. The experimental findings show the better performance of proposed model, achieving 99.12% accuracy, indicating its effectiveness for environmental monitoring.