Hi everyone — I'm KG, and I just completed Checkpoint 1 of the OpenAI to Z Challenge!
What began as a fevered idea — literally, while recovering from a high fever next to my quiet Shiba Inu named Fuwa — evolved into an actual geospatial analysis pipeline powered by GPT-4, Google Earth Engine, and NDVI data.
🌿 The Challenge
This global competition invites participants to uncover lost or hidden archaeological patterns in the Amazon rainforest using satellite data and AI. The rules are open-ended, and the tools are up to you.
For Checkpoint 1, I set out to find five locations in Brazil where vegetation appears "unnatural" — meaning too geometric, too flat, or oddly shaped for untouched forest. These could indicate past or present human intervention, and perhaps even hidden ruins.
🛰️ My Approach
With GPT-4 as my guide, I combined:
GEDI LiDAR data (from NASA)
NDVI vegetation index from Landsat imagery
Deforestation polygons from TerraBrasilis
Python + Earth Engine API
Visual overlays (RGB composite + NDVI)
We searched for anomalies near water margins — areas legally protected but often subtly encroached upon. GPT helped not only with code, but also with interpreting patterns, checking assumptions, and even creating metaphors (like "GPU arcade gems") to reframe ideas.
📍 What We Found
We discovered 5 strange regions:
Rectangular cutouts deep in the forest
Grid-pattern farms in unexpected terrain
Circular irrigation plots in the Cerrado
Fishbone road layouts
Cleared patches hugging rivers too closely
Each location was mapped with NDVI and RGB overlays, then cross-checked with deforestation records.
🐶 Fuwa’s Role
Fuwa — a calm, fluffy Shiba Inu — was by my side the whole time.
While I chased pixels and cloud masks, she stared silently at the screen, as if sensing something deeper.
Her stillness reminded me that sometimes, the best discoveries come not from more data, but from watching how nature should feel — and noticing when something’s off.
🧠 GPT-4: Not Just a Tool
I treated GPT-4 not like a calculator, but like a co-researcher.
It questioned me. It debugged me. It reminded me that sometimes, geometry speaks louder than words — especially in the jungle.
✅ Checkpoint 1: Done!
I’ve just posted my results in the official Discord.
It includes:
5 location maps
NDVI + RGB overlays
Annotated explanations
Code samples
PDF summary
I'm now working on an interactive dashboard, and thinking of training a small anomaly-detection model for follow-up exploration.
☕ If You Enjoyed This…
This took many late nights, satellite downloads, and gentle barks of encouragement.
If you'd like to support my exploration — or just fuel the next rabbit hole — you can buy me a coffee ☕ below.
Let’s see where the jungle leads.
— KG & Fuwa
Osaka, Japan
