🌊 CoastalAI

Advanced AI/ML Learning Platform for Coastal Oceanography

Explore how artificial intelligence and machine learning revolutionize our understanding of physical and biogeochemical processes in coastal marine environments by Claudio Iturra RL

🎓 Learning Objectives

Understanding AI/ML applications in coastal oceanographic research and monitoring

🌊 Physical Processes

12+

Key physical oceanographic processes including currents, waves, tides, mixing, and coastal upwelling

🧬 Biogeochemical Cycles

8+

Essential biogeochemical processes: nutrient cycling, primary production, carbon dynamics, and oxygen variability

🤖 AI/ML Techniques

15+

Machine learning methods: neural networks, time series analysis, computer vision, and ensemble methods

📈 Real Applications

25+

Practical case studies from coastal monitoring, climate research, and ecosystem management

🎯 What You'll Learn

Physical Oceanography + AI

  • 🌊 Current prediction using LSTM networks
  • 🌀 Eddy detection with computer vision
  • 📊 Wave forecasting with ensemble methods
  • 🌡️ Temperature field reconstruction
  • ⚡ Real-time coastal monitoring systems

Biogeochemical Processes + AI

  • 🦠 Phytoplankton bloom prediction
  • 💨 Carbon flux estimation
  • 🧪 Nutrient cycle modeling
  • 🐟 Ecosystem health assessment
  • 🔬 Water quality forecasting

📡 Data Collection

Sensors, satellites, autonomous vehicles

🔧 Preprocessing

Quality control, gap filling, normalization

🧠 AI/ML Analysis

Pattern recognition, prediction, classification

🎯 Applications

Forecasting, monitoring, decision support

🌊 Physical Oceanography & AI

Advanced methods for understanding coastal physical processes using machine learning

🌊 Coastal Current Systems

Geostrophic Current Calculation

Geostrophic Balance:
u_g = -(g/f) × (∂η/∂y)
v_g = (g/f) × (∂η/∂x)
Where: g = gravitational acceleration, f = Coriolis parameter, η = sea surface height

Thermal Wind Relation

Vertical Shear:
∂u_g/∂z = -(g/f×ρ₀) × (∂ρ/∂y)
∂v_g/∂z = (g/f×ρ₀) × (∂ρ/∂x)
Links horizontal density gradients to vertical current shear

🤖 AI Applications for Current Analysis

  • LSTM Networks: Predict current velocities from historical data and meteorological forcing
  • Convolutional Neural Networks: Extract current patterns from satellite altimetry and SST images
  • Gaussian Process Regression: Interpolate sparse current measurements with uncertainty quantification
  • Ensemble Methods: Combine multiple models for robust current forecasting

🌀 Mesoscale Eddies and Circulation

Vorticity Dynamics

Relative Vorticity:
ζ = ∂v/∂x - ∂u/∂y

Potential Vorticity:
PV = (f + ζ)/h
Where: f = planetary vorticity, ζ = relative vorticity, h = layer thickness

Okubo-Weiss Parameter

Eddy Detection Criterion:
W = S² - ζ²
S² = (∂u/∂x - ∂v/∂y)² + (∂v/∂x + ∂u/∂y)²
W < 0 indicates rotational dominance (eddy cores)

🤖 AI for Eddy Detection and Tracking

  • Computer Vision: Automated eddy identification in satellite imagery using U-Net architectures
  • Object Tracking: Kalman filters and particle filters for eddy trajectory prediction
  • Deep Learning: ResNet models for eddy classification (cyclonic vs anticyclonic)
  • Time Series Analysis: Detect eddy-induced variability in coastal measurements

🌊 Wave Dynamics and Coastal Processes

Linear Wave Theory

Dispersion Relation:
ω² = gk × tanh(kh)
Wave Celerity:
c = ω/k = √(g/k × tanh(kh))
Where: ω = angular frequency, k = wavenumber, h = water depth

Wave Energy and Momentum

Wave Energy Density:
E = (1/2)ρg⟨η²⟩
Wave Power:
P = E × c_g = (ρg²/16π) × H²T
Where: H = significant wave height, T = wave period, c_g = group velocity

🤖 AI for Wave Prediction and Analysis

  • Neural Networks: Wave height forecasting using meteorological inputs and historical wave data
  • Spectral Analysis: ML-enhanced wave spectrum estimation from buoy measurements
  • Image Processing: Wave parameter extraction from coastal video monitoring systems
  • Ensemble Forecasting: Probabilistic wave predictions for coastal engineering applications

🌡️ Temperature and Stratification

Heat Budget Equation

Temperature Evolution:
∂T/∂t + u⃗·∇T = ∇·(K∇T) + Q/(ρc_p)
Where: K = diffusivity tensor, Q = heat sources/sinks, c_p = specific heat

Brunt-Väisälä Frequency

Stratification Strength:
N² = -(g/ρ₀) × (∂ρ/∂z)
N² ≈ (g/ρ₀) × (α∂T/∂z - β∂S/∂z)
Where: α = thermal expansion coefficient, β = haline contraction coefficient

🤖 AI for Temperature Field Analysis

  • Data Assimilation: Optimal interpolation of temperature profiles using machine learning
  • Pattern Recognition: Identify thermal fronts and upwelling events in satellite SST data
  • Predictive Modeling: Forecast temperature stratification changes using LSTM networks
  • Anomaly Detection: Identify unusual thermal events and marine heatwaves

🌊 Coastal Upwelling and Mixing

Ekman Transport

Wind-Driven Transport:
M_x = τ_y/(ρf), M_y = -τ_x/(ρf)
Upwelling Velocity:
w = (1/ρf) × (∂τ_x/∂y - ∂τ_y/∂x)
Where: τ = wind stress, f = Coriolis parameter

Turbulent Kinetic Energy

TKE Budget:
∂k/∂t + u⃗·∇k = P + B - ε + ∇·(K_k∇k)
Where: P = shear production, B = buoyancy production, ε = dissipation

🤖 AI for Upwelling and Mixing Analysis

  • Feature Detection: Identify upwelling signatures in multi-sensor satellite data
  • Process Modeling: ML-enhanced parameterizations for mixing processes
  • Forecasting: Predict upwelling intensity and timing using atmospheric forcing
  • Data Fusion: Combine multiple data sources for comprehensive upwelling monitoring

🧬 Biogeochemical Oceanography & AI

Machine learning approaches to understanding coastal biogeochemical processes and ecosystem dynamics

🦠 Primary Production and Phytoplankton Dynamics

Photosynthesis and Growth

Photosynthesis-Irradiance Relationship:
P = P_max × (1 - e^(-αI/P_max))
Phytoplankton Growth Rate:
μ = μ_max × f(I) × f(N) × f(T)
Where: P_max = maximum photosynthetic rate, α = photosynthetic efficiency, I = irradiance

Nutrient Limitation (Monod Kinetics)

Nutrient Uptake:
f(N) = N/(K_s + N)
Multiple Nutrient Limitation:
f(N₁,N₂,...) = min(N₁/(K₁+N₁), N₂/(K₂+N₂), ...)
Where: K_s = half-saturation constant for nutrient uptake

🤖 AI for Primary Production Analysis

  • Remote Sensing ML: Estimate chlorophyll-a and primary production from satellite ocean color using neural networks
  • Bloom Prediction: LSTM models for harmful algal bloom forecasting using environmental drivers
  • Species Classification: Deep learning for phytoplankton species identification from microscopy images
  • Ecosystem Modeling: ML-enhanced biogeochemical models with adaptive parameterizations

💨 Carbon Cycle and Air-Sea Exchange

Carbonate System

CO₂ Equilibrium:
CO₂ + H₂O ⇌ H₂CO₃ ⇌ H⁺ + HCO₃⁻ ⇌ 2H⁺ + CO₃²⁻
pH Calculation:
pH = -log₁₀[H⁺]
[H⁺] = K₁[CO₂]/[HCO₃⁻] = K₂[HCO₃⁻]/[CO₃²⁻]

Air-Sea CO₂ Flux

Gas Exchange:
F = k × K₀ × (pCO₂_air - pCO₂_sea)
Gas Transfer Velocity:
k = k₀ × (Sc/660)^(-0.5) × f(u₁₀)
Where: K₀ = solubility, Sc = Schmidt number, u₁₀ = wind speed at 10m

🤖 AI for Carbon Cycle Analysis

  • pCO₂ Mapping: Neural networks for spatial interpolation of sparse pCO₂ measurements
  • Flux Estimation: Machine learning models for air-sea CO₂ flux using satellite and reanalysis data
  • Ocean Acidification: Predictive models for pH and aragonite saturation state
  • Carbon Budget: AI-assisted carbon accounting in coastal ecosystems

🧪 Nutrient Cycling and Biogeochemical Processes

Nitrogen Cycle

Nitrification:
NH₄⁺ + O₂ → NO₂⁻ + H₂O + H⁺ → NO₃⁻ + H₂O
Denitrification:
NO₃⁻ → NO₂⁻ → NO → N₂O → N₂
N:P Redfield Ratio:
C:N:P = 106:16:1 (molar ratio)

Phosphorus and Silicon Cycles

Phosphate Regeneration:
Rate = k_reg × [POM] × f(T) × f(O₂)
Silicate Dissolution:
Rate = k_diss × [BSi] × (1 - [Si(OH)₄]/[Si(OH)₄]_sat)
Where: POM = particulate organic matter, BSi = biogenic silica

🤖 AI for Nutrient Cycle Analysis

  • Nutrient Mapping: Gaussian process regression for nutrient field reconstruction
  • Process Rates: ML estimation of biogeochemical transformation rates
  • Stoichiometry: AI detection of deviations from Redfield ratios
  • Seasonal Cycles: Time series analysis of nutrient variability patterns

🫁 Oxygen Dynamics and Hypoxia

Oxygen Budget

Oxygen Balance:
∂O₂/∂t = P - R - ∇·(u⃗O₂) + ∇·(K∇O₂) + F_atm
Saturation Concentration:
O₂_sat = f(T,S,P) (Garcia & Gordon, 1992)
Where: P = production, R = respiration, F_atm = air-sea exchange

Apparent Oxygen Utilization

AOU Calculation:
AOU = O₂_sat - O₂_observed
Oxygen Utilization Rate:
OUR = AOU / τ (where τ = water mass age)

🤖 AI for Oxygen and Hypoxia Analysis

  • Hypoxia Prediction: Early warning systems using ML models and environmental indicators
  • Dead Zone Mapping: Computer vision for hypoxic zone detection in satellite data
  • Metabolic Rates: AI estimation of community respiration and production rates
  • Ecosystem Health: Machine learning indicators of oxygen stress in marine ecosystems

🐟 Marine Ecosystem Dynamics

Food Web Structure

Lotka-Volterra Equations:
dN₁/dt = r₁N₁ - α₁₂N₁N₂
dN₂/dt = -r₂N₂ + α₂₁N₁N₂
Holling Type II Functional Response:
f(N) = aN/(1 + ahN)
Where: r = intrinsic growth rate, α = interaction coefficient, a = attack rate, h = handling time

Biodiversity Indices

Shannon Diversity:
H' = -Σ(p_i × ln(p_i))
Simpson's Index:
D = Σ(p_i²), where p_i = proportion of species i

🤖 AI for Ecosystem Analysis

  • Species Distribution: MaxEnt and ensemble models for habitat suitability mapping
  • Biodiversity Assessment: Deep learning for automated species counting and identification
  • Ecosystem Indicators: ML-derived indices of ecosystem health and stability
  • Fisheries Management: AI-assisted stock assessment and catch prediction models

🤖 AI/ML Methods for Oceanography

Comprehensive overview of machine learning techniques applied to coastal oceanographic research

🧠 Deep Learning Architectures

🔗 Recurrent Neural Networks (RNNs)

LSTM Cell:
f_t = σ(W_f · [h_{t-1}, x_t] + b_f)
i_t = σ(W_i · [h_{t-1}, x_t] + b_i)
C̃_t = tanh(W_C · [h_{t-1}, x_t] + b_C)
C_t = f_t * C_{t-1} + i_t * C̃_t

Applications: Time series forecasting, current prediction, biogeochemical cycle modeling

🖼️ Convolutional Neural Networks (CNNs)

Convolution Operation:
(f * g)(t) = ∫ f(τ)g(t - τ)dτ
Feature Map:
Y_{i,j} = σ(Σ Σ W_{m,n} × X_{i+m,j+n} + b)

Applications: Satellite image analysis, eddy detection, species classification from imagery

🎯 Attention Mechanisms

Attention Weights:
α_i = exp(e_i) / Σ exp(e_j)
Context Vector:
c = Σ α_i × h_i

Applications: Multi-variate time series analysis, spatial-temporal pattern recognition

🔄 Autoencoders

Encoder:
z = f(Wx + b)
Decoder:
x' = g(W'z + b')

Applications: Dimensionality reduction, anomaly detection, data compression

📊 Statistical Learning Methods

🌳 Random Forest

Ensemble Prediction:
ŷ = (1/B) Σ T_b(x)
Out-of-Bag Error:
OOB = (1/n) Σ L(y_i, ŷ_{-i})

Applications: Variable importance analysis, non-linear relationship modeling

📈 Gaussian Process Regression

Posterior Mean:
μ(x*) = k(x*)ᵀ(K + σ²I)⁻¹y
Posterior Variance:
σ²(x*) = k(x*,x*) - k(x*)ᵀ(K + σ²I)⁻¹k(x*)

Applications: Spatial interpolation with uncertainty, optimal sensor placement

🎯 Support Vector Machines

Decision Function:
f(x) = Σ α_i y_i K(x_i, x) + b
RBF Kernel:
K(x_i, x_j) = exp(-γ||x_i - x_j||²)

Applications: Classification of water masses, pattern recognition in complex datasets

🔍 Principal Component Analysis

Eigenvalue Decomposition:
C = XᵀX = VΛVᵀ
Principal Components:
Y = XV (dimensionality reduction)

Applications: EOF analysis, mode decomposition, data visualization

⚡ Specialized Techniques

🌊 Physics-Informed Neural Networks (PINNs)

Loss Function:
L = L_data + λ_pde × L_pde + λ_bc × L_bc
PDE Residual:
L_pde = ||∂u/∂t + N[u] - f||²

Applications: Solving PDEs with sparse data, incorporating physical constraints

🔄 Generative Adversarial Networks (GANs)

Minimax Game:
min_G max_D V(D,G) = E[log D(x)] + E[log(1-D(G(z)))]

Applications: Synthetic data generation, super-resolution of satellite imagery

🎲 Bayesian Neural Networks

Posterior Distribution:
p(w|D) = p(D|w)p(w) / p(D)
Predictive Distribution:
p(y*|x*,D) = ∫ p(y*|x*,w)p(w|D)dw

Applications: Uncertainty quantification, robust predictions with confidence intervals

🔗 Graph Neural Networks

Message Passing:
m_{ij} = M(h_i, h_j, e_{ij})
Node Update:
h_i' = U(h_i, Σ m_{ij})

Applications: Modeling connectivity in marine ecosystems, sensor network analysis

🎯 Key Considerations for Oceanographic AI

  • Data Quality: Handle missing data, sensor drift, and measurement uncertainties
  • Temporal Scales: Account for multiple time scales from tidal to seasonal to interannual
  • Spatial Heterogeneity: Consider coastal complexity and boundary effects
  • Physical Constraints: Incorporate conservation laws and known physical relationships
  • Interpretability: Ensure models provide insights into underlying processes

📊 Real-World Case Studies

Practical applications of AI/ML in coastal oceanographic research and monitoring

🌊 Case Study 1: Gulf Stream Eddy Prediction

🎯 Objective

Develop an AI system to predict Gulf Stream warm-core ring formation and propagation for fisheries management and navigation safety.

📊 Data Sources

  • AVISO satellite altimetry (1993-present)
  • MODIS sea surface temperature
  • HYCOM model outputs
  • Argo float profiles

🤖 ML Approach

  • Detection: U-Net CNN for eddy identification
  • Tracking: Kalman filter with LSTM state prediction
  • Forecasting: Ensemble of LSTM networks

📈 Results

Detection Accuracy

94.2%

7-day Forecast Skill

0.78

Position Error (km)

15.3

🔬 Key Findings

  • CNN outperformed traditional geometric methods
  • Multi-sensor fusion improved tracking accuracy
  • Ensemble approach reduced forecast uncertainty
  • Real-time implementation achieved 2-hour latency

🦠 Case Study 2: Harmful Algal Bloom Early Warning

🎯 Objective

Create a predictive system for toxic Alexandrium blooms in the Gulf of Maine to protect public health and shellfish industry.

📊 Data Integration

  • MODIS ocean color (chlorophyll, turbidity)
  • NOAA weather and oceanographic buoys
  • Shellfish toxicity monitoring data
  • River discharge and nutrient loading
  • Alexandrium cyst abundance surveys

🧠 AI Architecture

  • Feature Engineering: Environmental indices and lag variables
  • Classification: Random Forest for bloom/no-bloom prediction
  • Regression: XGBoost for toxin concentration
  • Ensemble: Weighted combination of multiple models

📊 Performance Metrics

Precision

87%

Recall

92%

Lead Time (days)

14

💡 Impact

  • Reduced false alarms by 40%
  • Extended forecast lead time from 3 to 14 days
  • Saved $2.3M annually in avoided closures
  • Integrated into NOAA operational forecasting

🌡️ Case Study 3: Marine Heatwave Prediction

🎯 Challenge

Predict marine heatwave onset, intensity, and duration in the California Current System for ecosystem management and fisheries adaptation.

🔄 Methodology

Marine Heatwave Definition:
MHW when: SST > 90th percentile for ≥5 consecutive days
Intensity Categories:
I_mean = (T_mean - T_90) / (T_90 - T_mean)
Moderate: 1-2×, Strong: 2-3×, Severe: 3-4×, Extreme: >4×

🤖 Deep Learning Pipeline

  • Data: 40 years of OISST, atmospheric reanalysis
  • Architecture: ConvLSTM for spatiotemporal patterns
  • Training: Transfer learning from global to regional
  • Validation: Cross-validation with recent events

🎯 Forecast Performance

30-day Skill Score

0.65

Duration RMSE (days)

8.2

Intensity Correlation

0.73

🐟 Ecological Applications

  • Salmon migration timing adjustments
  • Kelp forest vulnerability assessment
  • Fisheries quota management
  • Marine protected area effectiveness

💨 Case Study 4: Coastal Carbon Flux Monitoring

🎯 Research Goal

Quantify air-sea CO₂ exchange in coastal waters using machine learning to fill gaps in sparse observations and improve carbon budget estimates.

📡 Observational Network

  • Autonomous surface vehicles (pCO₂)
  • Coastal buoy measurements
  • Ship-based surveys
  • Satellite wind and SST
  • Biogeochemical Argo floats

🧮 ML Framework

CO₂ Flux Equation:
F = k × K₀ × (pCO₂_air - pCO₂_sea)
ML Enhancement:
pCO₂_sea = f_ML(SST, SSS, Chl, Wind, DOY, Lat, Lon)

📈 Model Performance

pCO₂ RMSE (μatm)

18.5

Flux R² Correlation

0.81

Uncertainty (%)

±15

🌍 Climate Impact

  • Revised coastal carbon sink estimates
  • Improved global carbon cycle models
  • Policy support for blue carbon initiatives
  • Enhanced climate change projections

🔑 Success Factors Across Case Studies

📊 Data Strategy

  • Multi-sensor data fusion
  • Long-term historical records
  • Quality control and validation
  • Real-time data streams

🤖 Model Design

  • Domain-specific architectures
  • Ensemble approaches
  • Uncertainty quantification
  • Interpretable features

🔄 Operational Integration

  • Real-time processing pipelines
  • User-friendly interfaces
  • Stakeholder engagement
  • Continuous model updates

📈 Impact Assessment

  • Quantified economic benefits
  • Scientific advancement metrics
  • Societal impact evaluation
  • Policy influence tracking

⚙️ Interactive Oceanographic AI Simulator

Hands-on experience with AI/ML methods applied to synthetic coastal oceanographic data

🌊 Coastal Process Simulator

90 days
10%
70%

📊 Simulation Results

Configure parameters above and click "Generate Data" to start the simulation...

🎯 Model Accuracy

--

R² correlation coefficient

📊 RMSE

--

Root mean square error

⚡ Training Time

--

Seconds

🔮 Forecast Skill

--

7-day prediction accuracy

📚 Understanding the Results

🎯 Model Performance Metrics

  • R² (Coefficient of Determination): Proportion of variance explained by the model (0-1, higher is better)
  • RMSE (Root Mean Square Error): Average prediction error in original units (lower is better)
  • MAE (Mean Absolute Error): Average absolute difference between predictions and observations
  • Forecast Skill: Ability to predict future values beyond climatology

🔍 Interpretation Guidelines

  • High Noise: More complex models may overfit; simpler models often perform better
  • Limited Data: Ensemble methods and regularization help prevent overfitting
  • Seasonal Patterns: LSTM networks excel at capturing temporal dependencies
  • Spatial Patterns: CNNs are effective for image-like oceanographic data