resources/RESOURCES_CHECKLIST.md

Resources Checklist

Central to-do hub for all data, models, and content needed before the portfolio app is ready for public deployment. Check off items as they are completed.

---

Configuration & Infrastructure

• GEE Service Account — create a service account in Google Cloud Console, grant it Earth Engine access, download the JSON key, and add to:
  • Local: .streamlit/secrets.toml (not committed to git)
  • Deployed: Streamlit Cloud secrets dashboard
  • Format:

    [gee]
    service_account_email = "my-account@my-project.iam.gserviceaccount.com"
    private_key = "-----BEGIN RSA PRIVATE KEY-----\n..."
    

• GitHub repo URL — update the st.link_button on Page 4 with the real URL
• streamlit-lottie installed: pip install streamlit-lottie>=0.0.5
• Add streamlit-folium and streamlit-lottie to pyproject.toml

---

Hero Animation (Page 1)

• Find a bear/forest Lottie animation at https://lottiefiles.com
  • Search: "bear", "forest", "tree", "nature", "fire prevention"
  • Download as JSON → save to resources/smokey_animation.json
  • Then uncomment the streamlit-lottie block in app/streamlit_app.py (Page 1)
  • Install: from streamlit_lottie import st_lottie
• Alt: Download a high-quality animated GIF → resources/smokey.gif and embed via st.image

---

Content To-Dos (author before public deploy)

Page 1 — "Why This Matters"

• Replace placeholder text with real framing:
  • International forest restoration commitments (e.g., scale of loss globally)
  • Why satellite-based detection is better than ground surveys
  • What makes this project novel or useful

Page 3 — "Why I Picked the Winner"

• Write real reasoning in the st.expander on Page 3:
  • Why PR-AUC over accuracy for 4.8% class imbalance
  • Brier score as a calibration check (not just ranking)
  • Recall vs. precision trade-off for environmental monitoring (false negatives = missed deforestation events = real-world consequence)
  • What the Optuna tuning added vs. the baseline XGBoost

Page 4 — Architecture Node Descriptions

• Fill in the Sankey node descriptions in the nodes dict (app/streamlit_app.py, ~line 700):
  • Google Earth Engine
  • Parquet Export
  • Feature Engineering
  • Optuna Tuning
  • XGBoost Model
  • FastAPI
  • Streamlit Dashboard

Page 4 — Key Decisions & Lessons

• Write real content in the 4 st.expander blocks:
  • "Why cosine drift as the core signal"
  • "Handling severe class imbalance (4.8%)"
  • "Why geo-stratified train/test split"
  • "Optuna over grid search"

---

AOI Fun Facts Research

Each AOI rectangle on the map (Pages 2 + 3) shows a fun fact on click. Replace the PLACEHOLDER entries in AOI_FUN_FACTS in app/streamlit_app.py.

• Indonesia-Malaysia — highest drift in dataset; research: peat fires, palm oil, logging in Borneo/Sumatra. Add source URL.

  Suggested sources: Global Forest Watch, Mongabay, Nature journal

• Amazon Basin — add a compelling fact + source
• Congo Basin DRC — add a compelling fact + source
• Guinea — add a compelling fact + source
• Canada — add a compelling fact + source
• Mekong Region — add a compelling fact + source
• Cerrado Brazil — add a compelling fact + source

---

Pre-Deploy Checklist

• All content to-dos above are filled in
• All pre-computed artifact boxes above are checked
• GEE service account is set up and tested (click a point on Page 3 map)
• GitHub link on Page 4 is updated
• Hero animation or image is in place
• App runs locally with streamlit run app/streamlit_app.py without errors
• All 4 pages load in < 2 seconds (no 800k-row parquet loads at runtime)

---

---

✅ Done

Pre-computed Data Artifacts

Run python resources/precompute.py to regenerate all artifacts.

• resources/kpi_summary.json — KPI cards on Page 1
• resources/mean_embedding_profile.parquet — embedding profile chart on Page 2
• resources/drift_by_area_year.parquet — drift analysis chart + violin on Page 2
• resources/target_distribution.json — class distribution charts on Page 2
• resources/timelapse_sample.parquet — timelapse animation + map dots on Page 2
• resources/canada_best_trial.json + amazon_basin_best_trial.json — model metrics on Page 3
• resources/canada_feature_importance.json + amazon_basin_feature_importance.json — Page 3 side-by-side
• resources/canada_confusion_matrix.json + amazon_basin_confusion_matrix.json — Page 3 side-by-side
• resources/canada_data_sample.parquet + amazon_basin_data_sample.parquet — Page 2 area toggle
• resources/comparison_kpis.json — "Tale of two forests" on Page 1
• resources/aoi_stats.json — sidebar Other Regions stats table
• resources/evaluations_clean.parquet — superseded by per-area artifacts; no longer loaded by app
• resources/feature_importance.json — superseded by per-area feature importance; no longer loaded
• resources/confusion_matrix.json — superseded by per-area confusion matrices; no longer loaded
• resources/pca_sample.parquet — PCA scatter removed from app; no longer loaded

Model Files

• models/xgboost_canada_optuna.pkl and models/xgboost_amazon_basin_optuna.pkl — final tuned winners — copied to resources/ for HF Space
• models/rf_tuned.pkl, models/sgd_optuna.pkl, models/baseline_logistic_regression_*.pkl — were used in old 4-model comparison table; that page has been replaced by the Canada vs Amazon Basin framing; no longer needed by the app

streamlit-folium installed

• streamlit-folium>=0.25.0 installed and in use

Nav Bar Icons (optional polish)

• SVG icons confirmed working — implementation commented out in app/streamlit_app.py. To re-enable:
  1. Source per-page SVG icons (e.g., Lordicon, Heroicons, Phosphor) and save to resources/
  2. Add import base64 to top-level imports in app/streamlit_app.py
  3. Uncomment _load_nav_icon() and the button loop in the comment block above _render_nav()
  4. Optionally remap SVG stroke colors to theme values (#40916c, #1b4332) via string replace before base64 encoding
  • Current placeholder: resources/wired-outline-955-demand-hover-click.svg (Lordicon "demand" icon, tested ✓)
  • Lottie animations in nav — not practical inside nav buttons (renders as full-height component block)

Theme Selection

• Theme selected and applied: Canopy Light + Manrope — THEME dict set in app/streamlit_app.py