ahmedoussemabouhageb/Tuna-Fish-Zone-Predection-ai

Deep learning ensemble (ConvLSTM + Autoencoder + Vision Transformer) for predicting tuna fishing hotspots off North Tunisia using daily ocean environmental data Tuna Fishing Hotspot Prediction North Tunisia A deep learning project for predicting tuna fishing hotspots in the waters off North Tunisia using ocean environmental data and historical fishing activity. The model ensemble combines ConvLSTM, a Convolutional Autoencoder, and a Vision Transformer to forecast where fishing vessels are most likely to concentrate during peak tuna season. Overview Tuna fishing activity in the Mediterranean is highly seasonal and spatially concentrated. This project learns to predict the most likely fishing hotspots by correlating historical vessel activity with environmental ocean conditions such as sea temperature, salinity, currents, and chlorophyll concentration. Rather than predicting exact fishing hours, the model ranks spatial cells by likelihood of activity, making it useful for planning and resource allocation. The pipeline covers vessel data ingestion, ocean feature extraction, hotspot identification, model training across three architectures, and ensemble evaluation with test-time augmentation. Study Area Region: North Tunisia coastal waters Latitude: approximately 38.0°N to 38.75°N Longitude: approximately 8.0°E to 9.5°E Peak season: May, June, and July (months with historically significant tuna catch) Spatial resolution: 0.25° grid Data Sources Vessel Activity (Global Fishing Watch) Daily fishing effort records from 2019 to 2024 Columns used: latitude, longitude, date, fishing hours, vessel hours, vessel count (MMSI), and gear type Gear types weighted by fishing intensity: tuna purse seines weighted highest, drifting longlines as baseline Ocean Environmental Data (Copernicus Marine Service) Sea surface currents (U and V components) Sea water temperature Salinity Chlorophyll-a concentration Sea surface height (SSH / zos) Mixed layer depth (MLD) Bathymetry (static, log-transformed) All ocean data is filtered to peak months only and interpolated to a common spatial grid. Hotspot Detection Before training, …