Skip to content

CLAUDE.md

Latest commit

 

History

History
83 lines (56 loc) · 3.65 KB

CLAUDE.md

File metadata and controls

83 lines (56 loc) · 3.65 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

UT Austin Geoscience Hackathon project by Team Alpha. Analyzes land use and environmental change using AlphaEarth satellite embeddings (64-dimensional AI vectors) and LandTrendr (traditional NDVI-based spectral-temporal change detection), comparing both methods across 15 study sites in the western United States.

Environment Setup

No requirements file exists. Key dependencies (install via pip or conda):

pip install earthengine-api geemap numpy pandas matplotlib seaborn scipy ipyleaflet ipywidgets

Authenticate with Google Earth Engine before running notebooks:

earthengine authenticate

All notebooks require an active GEE session (ee.Initialize()).

Notebook Architecture

The project is entirely notebook-based — no standalone scripts are used in the main workflow.

Main Notebooks (run in this order for full analysis):

Notebook Purpose
alpha_tutorial.ipynb Introductory tutorial: AlphaEarth K-means clustering
AlphaEarth_Story.ipynb Primary narrative: cosine similarity, dam detection, Austin urban growth
AlphaEarth_Interactive_Mapping.ipynb Interactive AOI drawing, real-time change detection, export tools
AlphaEarth_LandTrendr_ChangeComparison.ipynb Side-by-side comparison of both algorithms across 15 sites + IoU analysis
LandTrendr_AlphaEarth.ipynb Deep-dive LandTrendr analysis, statistical summaries, CSV exports

Backup notebooks in /backup/:

  • AlphaEarth_EDA.ipynb, LandTrendr_EDA.ipynb — exploratory/tuning work
  • Austin_LandTrendr_Changes.ipynb — Austin-specific analysis
  • Austin_LandTrendr.py — standalone script version

Key Data Sources (Google Earth Engine)

  • AlphaEarth embeddings: GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL — 64-band annual composite, 2017–2024
  • LandTrendr outputs: 15 pre-computed GEE assets (one per study site), 2016–2024 NDVI-based
  • Landsat: Used internally by LandTrendr for NDVI computation

Change Detection Methodology

Both algorithms output three standardized layers:

  • YOD — Year of first detected change
  • MAG — Magnitude of change
  • DUR — Duration (consecutive high-change years)

AlphaEarth pipeline: Cosine similarity between embedding years → Gaussian smoothing (kernel radius=2, sigma=1) → threshold >0.15 → morphological operations

LandTrendr pipeline: NDVI spectral trajectory fitting → threshold >170 NDVI units → same morphological cleanup

Validation: IoU (Intersection over Union) computed between both methods' binary change masks per site.

Study Sites

Urbanization: Austin TX, Dallas TX, Houston TX, Bend OR, Portland OR, Sacramento CA Wildfires: Bootleg OR (2021), Camp Fire CA (2018), Dixie CA (2021), Mosquito CA (2022), Santiam OR (2020) Forest/Logging: Angelina TX, Coos Bay OR, Mt Hood OR, Shasta-Trinity CA

Interactive Features

AlphaEarth_Interactive_Mapping.ipynb provides:

  • Draw AOI on map → auto-computes change layers server-side in GEE
  • Inspector tool: click any point to see year-by-year similarity chart
  • Sliders: magnitude threshold, smoothing radius/sigma, layer opacity
  • Export to GEE Asset or Google Drive

Output Files

  • lt_training_stats_all_assets.csv — per-band stats for all 15 LandTrendr assets
  • lt_training_corr_all_assets.csv — inter-band correlations
  • lt_training_asset_summary.csv — training pixel counts and areas
  • images/IoU_LandTrendr_AlphaEarth_15sites.png — comparison result figure
  • UT_GIS_map/ — exported map PNGs for Austin/UT campus change layers