Class

Main Theme

Theoretical Part (45-50 mins)

Practical Hands-on Part (70-75 mins)

1

GeoAI Python Stack & Vector Data Mastery

Introduction to GeoAI, course structure. Vector data models (points, lines, polygons). CRS fundamentals. GeoPandas, Shapely, Fiona ecosystem.

Hands-on: Load city shapefile (Mumbai wards). Perform spatial joins (schools per ward). Reproject CRS. Visualize with matplotlib.

2

Cloud-Native Rasters & Interactive Dashboards

Raster data (pixels, bands, resolution). Cloud-Optimized GeoTIFFs (COGs), STAC. Rasterio & xarray. Principles of effective geospatial visualization.

Hands-on: Clip DEM to city boundary. Calculate slope/aspect. Use stackstac to load Sentinel-2 from Planetary Computer. Build Streamlit + Folium dashboard with NDVI time-series slider.

3

Feature Engineering & Crop Yield Prediction

ML workflow for geospatial data. Feature extraction from rasters: NDVI, LAI, EVI. Spatial cross-validation. Green/Blue/Grey Water Footprint concepts.

Hands-on: Extract Sentinel-2 NDVI time-series & weather data for agricultural fields. Engineer cumulative NDVI & temperature anomalies. Train XGBoost for crop yield. Calculate Green Water Footprint using SSEBop ET data.

4

Land Cover Classification & Water Footprint Integration

Classification algorithms (Random Forest). Blue Water Footprint: irrigation consumption from surface/groundwater. Remote sensing of irrigated areas.

Hands-on: Land use/land cover classification to identify irrigated cropland. Estimate Blue Water Footprint by combining irrigation data with soil moisture deficits.

5

Socio-Economic Vulnerability & Spatial Econometrics

Integrating census data with spatial data. Tobler’s First Law, spatial weights matrices, spatial lag/error models (spreg, libpysal).

Hands-on: Build multi-dimensional vulnerability index for flood-prone district (census + nightlights + infrastructure). Apply spatial lag model to understand poverty drivers.

6

Flood & Landslide Susceptibility with Explainable ML

Ensemble ML for hazard mapping: Random Forest vs. XGBoost. Topographic, hydrological, and geological conditioning factors. Hyperparameter tuning with GridSearchCV. Model evaluation: AUC‑ROC, precision‑recall for rare events. Introduction to Explainable AI: SHAP for feature attribution.

Hands-on: Train XGBoost model for Mumbai flood susceptibility using conditioning factors (elevation, slope, distance to drainage). Generate SHAP summary and dependence plots to identify primary flood drivers. Modify the pipeline for a landslide inventory (e.g., Himalayan foothills) and interpret feature importance differences.

7

Glacial Lake Outburst Flood (GLOF) Susceptibility & Impact

GLOF theory: lake formation, moraine dam stability, triggers. SAR remote sensing for lake mapping: backscatter characteristics of water/snow/ice. Sentinel‑1 preprocessing for change detection. Classification of lake stability (stable vs. potentially dangerous). Downstream impact zone simulation concepts (empirical volume‑area scaling).

Use Sentinel‑1 SAR data in Google Earth Engine or Python to delineate glacial lake boundaries in the Himalayas. Train a Random Forest classifier on morphological/geomorphometric lake features (area change, slope, distance to glacier). Overlay derived lake polygons with a simple DEM‑based flow‑routing model to visualize potential downstream inundation corridor (conceptual simulation).

8

Deep Learning for Earth Observation – U-Net & PyTorch

Fundamentals of CNNs. PyTorch tensors & dataloaders. U-Net architecture for semantic segmentation of satellite imagery.

Set up PyTorch dataloader for Sentinel-2. Load pre-trained U-Net to segment building footprints or water bodies.

9

Blue Carbon & Hyperspectral Imaging for Mineral Exploration

Mangrove/seagrass mapping with optical+SAR. Above-ground biomass estimation. Hyperspectral data (AVIRIS/PRISMA), spectral signatures of minerals. Spectral Angle Mapper (SAM).

Train U-Net to map Sundarbans mangroves with Sentinel-2. Estimate carbon stock. Use spectral library to identify mineral indicators in AVIRIS scene. Implement SAM classifier. Apply SHAP to CNN for band importance.

10

Physics-Informed ML for Rainfall-Runoff & Groundwater

Limitations of pure data-driven models. PINNs & torchphysics. Embedding mass balance into loss function. MODFLOW & FloPy introduction. Coupling PINN with MODFLOW for parameter estimation.

Build simple PINN for rainfall-runoff prediction in small catchment. Compare with XGBoost. Set up MODFLOW model with FloPy. Use PINN to estimate hydraulic conductivity from sparse well data.

11

Integrated Water Footprint & Scarcity Mapping

Green, Blue, Grey Water Footprint synthesis. Grey Water Footprint: dilution needs for nitrogen from fertilizers.

Calculate all three components for agricultural region. Estimate nitrogen loading from crop maps. Map total water footprint & identify scarcity hotspots against basin availability.

12

Urban Microclimate, Precision Agriculture & Marine Monitoring

Landsat thermal bands for LST. UHI modelling with ML & OSM data. Variable rate fertilization/prescription maps. Marine GeoAI: SST, chlorophyll-a, potential fishing zones.

Generate UHI map for Delhi (Landsat 8 + osmnx). Train model to predict UHI intensity. Create nitrogen prescription map from NDVI. Process SST & chlorophyll-a for Bay of Bengal; Random Forest for fishing zones; Streamlit dashboard.

13

Explainable AI (SHAP/LIME) & Causal GeoAI

Advanced SHAP & LIME applications. Global vs. local explanations. XAI for regulatory compliance & stakeholder trust. Causal reasoning, DAGs, uncertainty propagation.

Apply SHAP to crop yield model. Present feature importance plots. Compare predictive accuracy vs. decision confidence in hazard/crop models.

14

SAR Data & Time-Series Transformers

Synthetic Aperture Radar (SAR): backscatter, coherence, polarimetry. Pre-processing (calibration, speckle filter, terrain correction). Transformers for satellite time-series (crop classification, anomaly detection).

Map flood extent during monsoon with Sentinel-1 (pre/post change detection). Identify obstructed roads for disaster response. Use pre-trained transformer (or build simple one with PyTorch) to classify crop types from Sentinel-2 time-series. Automate glacial lake extraction from Landsat time-series.

15

Spatial Econometrics & Geometric Deep Learning

SLM, SEM, spatial weights (Queen/Rook/distance). Graph theory for spatial networks. Graph Convolutional Networks (GCNs) for rivers/roads.

PySAL/spreg spatial lag model for groundwater depletion (well density, rainfall, agriculture). Extract river network from DEM; PyTorch Geometric GCN to predict nitrate concentrations along river.

16

Point Cloud Learning for LiDAR (Forestry & Coastal Systems)

LiDAR for biomass, canopy height, coastal topography. PointNet/PointNet++ & semantic segmentation.

Open3D LiDAR visualization. PointNet for ground/vegetation classification. Calculate canopy height & carbon stock (mangroves). Coastal beach profiles.

17

Big Geospatial Data – Dask, Xarray & STAC/Pangeo

Challenges of petabyte-scale data. Dask for parallel single-machine workflows. STAC catalogs, Pangeo cloud-native geoscience. Optical/Radar/LiDAR fusion.

dask-geopandas + xarray with dask: monthly mean LST for entire country (chunking & parallel). Query STAC API (Planetary Computer) for Sentinel-1 & -2; stackstac + xarray to fuse optical & SAR; cloud masking with pystac-client.

18

Distributed Computing with PySpark (GeoSpark/Sedona)

Apache Spark, GeoSpark (Sedona), Spark GIS. Dask vs. Spark. Large-scale raster inference & vector joins.

Local PySpark + Sedona. Large-scale spatial join (millions of GPS points to boundaries). Distributed inference for land cover classification across tiles.

19

IoT Architecture, Protocols & Data Ingestion for Environmental Hazards

End‑to‑end IoT system architecture: sensors, gateway, cloud broker, application. Multi‑hazard node design considerations (flood, landslide, GLOF, wildfire). Communication protocols comparison: LoRaWAN (range vs. bandwidth), NB‑IoT, 4G/5G. REST API design principles for IoT data ingestion. Introduction to MQTT for lightweight telemetry.

Write a Python script to generate synthetic water level data (5‑min intervals) for a simulated flood sensor network. Use paho‑mqtt to publish the simulated data to a local MQTT broker (Mosquitto). Build a FastAPI endpoint (POST /sensor/reading) with Pydantic validation to receive and store the telemetry data in a lightweight database (e.g., SQLite or Redis). Write a simple Python script that acts as a mobile app sending POST requests to the API.

20

Real‑Time Dashboards & Early Warning System (EWS) Integration

Principles of effective real‑time geospatial dashboards. Connecting live data streams to web visualizations (WebSocket vs. polling). Alert level logic: threshold‑based vs. predictive alerts. Integrating static hazard susceptibility maps with dynamic sensor data for contextual early warning. User experience for emergency response dashboards.

Build a Streamlit + Folium dashboard that periodically fetches sensor readings from the FastAPI endpoint (built in Class 19). Display sensor markers with color‑coded alerts based on water level thresholds (e.g., green‑yellow‑red). Overlay the flood susceptibility map from Class 6A. When a sensor triggers a high alert, highlight the predicted inundation zone from the susceptibility map on the dashboard. Run the entire stack (API, dashboard) locally to simulate a working early warning prototype.

21

Introduction to Large Language Models (LLMs)

Introduction to Large Language Models (LLMs) and it's applications

-

22

Multimodal - Large Language Models (LLMs)

-

Hands-on: Build a multimodal LLM chatbot, Web-based geospatial apps, UI/UX for spatial data - Streamlit, Deck.gl

23

Introduction to Generative AI

Theory: Overview of generative models: GANs, VAEs, and diffusion models. Transformers, How GANs work (generator vs. discriminator). Applications in GeoAI: super‑resolution of satellite imagery, data augmentation for rare events (e.g., floods, landslides), generating synthetic land cover maps, and filling clouds in optical imagery.

-

24

GANs for Earth Observation

-

Hands-on: Build a simple GAN with PyTorch to generate synthetic NDVI time‑series for data‑scarce regions. Use a pre‑trained ESRGAN model to upscale Sentinel‑2 imagery from 10m to 2.5m resolution. Explore diffusion models for cloud removal on Landsat scenes.

25

Agentic AI

Introduction to Agents and GeoAI Agent

Hands-on: Build an Agent or Chat Bot

26

Multi-Agentic AI framework

Introduction to Multi Agentic Framework and orchestration Agent planning, execution, real-time monitoring/alerts, LangChain Agents, Collaborative roles, communication, task distribution. LangGraph

-Hands-on: Build a multiagentic framework for Climate/Agriculture Prediction and Forecasting

27

AI Orchestration

Chains, tools, memory management, stateful execution, LangChain, LangGraph

-

28

Containerization

The importance of containerization. Introduction to Docker.

Hands-on: Write a Dockerfile for the FastAPI app.

29

MLOps & API Dev

Introduction to MLOps. Model versioning with DVC or MLflow. CI/CD for GeoAI/ Building REST APIs with FastAPI. Input/output validation using pydantic.

Hands-on: Take the flood susceptibility model from Week 8. Package it into a FastAPI application. Create a /predict endpoint that accepts a GeoJSON point and returns the flood risk probability.

30

Cloud Deployment, Maintenance and Upscale

Introduction to Cloud Deployment and Its Importance

Deploying applications on cloud platforms (AWS/GCP). Using cloud storage (S3/GCS) and serverless functions (Lambda/Cloud Functions).

31

Introduction to Spatial RAG

What is RAG? Types of RAGs, Geospatial RAGs.

Hands-on: Use osmnx to create a vector database of building footprints with attributes. Use HuggingFace embeddings to create a retrieval system. Build a simple prompt that translates a user query like "find all parks within 500m of a school" into a geopandas query.

32

Capstone Project