How deusnexus/geo-ai-superres works
Project Type and Purpose
• This is a geospatial AI pipeline combining data ingestion, preprocessing, super‐resolution modeling, land‐cover classification, temporal analysis, and urban growth prediction.
• It operates as a research/analysis tool rather than a traditional deployed web service. Core interactions happen in Jupyter notebooks invoking modular Python code.
Key Components
a. Data Layer
• Raw satellite imagery (“Sentinel‐1,” “Sentinel‐2,” DEM) and auxiliary masks (NDVI, slope, water bodies) under data/raw and data/masks.
• Processed datasets (cloud‐masked, normalized) under data/processed.
• Super‐resolved outputs stored in data/superres.
• Time‐series data for trend analysis under data/temporal.
b. Processing Layer (src/data_processing)
• Data ingestion and chunking into patches.
• GeoJSON metadata handling for patches_on_land, highlighted_patches.
• Preprocessing steps called by notebooks (normalization, masking).
c. Modeling Layer (src/models and models/)
• Super‐resolution model (EHNet: CNN + Swin Transformer)
• Land‐cover classification model
• Urban growth prediction model
• Checkpoints and weights stored in models/*.pth
d. Analysis & Visualization (src/visualization, notebooks)
• Jupyter notebooks orchestrate entire pipeline:
– 00_* for data exploration and patch extraction
– 01_* to 05_* for training superres, classification, temporal analysis, growth prediction
• Visualization utilities for maps, charts, and generation of final reports and figures.
e. Outputs
• Final maps, charts, reports saved under outputs/maps, outputs/charts, outputs/reports.
Interactions and Data Flow
• Flow starts with raw imagery → data_processing modules → processed patches.
• Processed patches feed into model training pipelines for super‐resolution, land‐cover, and growth models.
• Trained models are loaded by notebooks for inference, generating high‐res outputs and predictions.
• Visualization modules consume predictions and raw/processed imagery to produce maps and dashboards.
Architectural Patterns and Design Principles
• Modular layering: clear separation between data ingestion, preprocessing, model training, and visualization.
• Notebook‐driven orchestration: notebooks as control plane that invoke Python modules in src/.
• Checkpointing: persistent model artifacts in models/ for reproducibility.
• Reusability & extensibility: src/utils for shared code; ability to swap models or add new analysis notebooks.