.env.example

# Copy to .env and uncomment values as needed (optional for local dev).
# python run.py sets APP_MODE=dda automatically.

# APP_MODE=dda
# REQUIRED for stable login sessions (including in production): without this,
# a random secret is generated per process start, invalidating every existing
# session/cookie on each restart. Generate one with:
#   python -c "import secrets; print(secrets.token_hex(32))"
# SECRET_KEY=change-me-to-a-long-random-string

# Database (default: SQLite in data/satellite_app.db)
# DATABASE_URL=sqlite:///./data/satellite_app.db
#
# MySQL (production for this project): install the driver first
# (pip install pymysql, already in requirements.txt), then set:
# DATABASE_URL=mysql+pymysql://user:password@host:3306/dbname
# URL-encode special characters in the password (e.g. @ -> %40).
#
# PostgreSQL also supported:
# DATABASE_URL=postgresql://user:pass@host/db

# Local image library folder (default: library_sources/)
# LOCAL_LIBRARY_ROOT=C:/path/to/your/images

# Image upload limit in MB (default 15360 = 15 GB)
# MAX_GEOTIFF_MB=15360
# MAX_IMAGE_MB=15360

# Automatic mode: identified same-area pairs can be queued on a cadence.
# The queue does NOT start when the app launches. On the Automatic tab, set
# interval (days) and run-at time (IST), then click Start. Manual mode is
# still click-to-run only. The local server must stay running while the
# schedule is armed. Set AUTO_DETECT_ENABLED=0 to hide/disable the feature.
# AUTO_DETECT_ENABLED=1
# AUTO_DETECT_INTERVAL_DAYS=10
# AUTO_DETECT_RUN_AT=02:00

# Max pixel dimension for detection (lower = less RAM, default 4096 local)
# DETECTION_MAX_SIDE=4096

# --- Accuracy controls (all optional; defaults preserve current behavior) ---
# Inference mode: downscaled | fullres_tiled | auto
# auto = fullres_tiled when before/after paths are GeoTIFF (RCA Critical #1)
# DETECTION_INFERENCE_MODE=auto
# Full-res cap (px) in fullres_tiled mode; 0 = native resolution
# DETECTION_FULLRES_MAX_SIDE=8192
# Native side (px) above which GeoTIFFs stream from disk windows (avoids OOM)
# Default lowered 8192→4096 (RCA Significant #5)
# DETECTION_WINDOWED_THRESHOLD=4096
# Soft RAM budget (MB) per in-memory array before switching to windowed streaming
# DETECTION_TILE_MEMORY_MB=1536
# Tile size / overlap for full-res scoring (overlap default 0.35 — RCA #6 seams)
# DETECTION_TILE_SIZE=512
# DETECTION_TILE_OVERLAP=0.35
# Skip SIFT registration on GPS-aligned GeoTIFF pairs: true|false|auto (RCA #4)
# DETECTION_SKIP_REGISTRATION_GEOTIFF=auto
# Local Delhi fine-tuned AdaptFormer (v3 frozen = Test F1 0.581 @ thr 0.2).
# Relative paths are resolved against the repo root (not the process CWD).
# ADAPTFORMER_WEIGHTS=models/adaptformer_delhi/v3_frozen
# Val-calibrated DL threshold (alias: DETECTION_DL_THRESHOLD). v3 = 0.2
# ADAPTFORMER_THRESHOLD=0.2
# DETECTION_DL_THRESHOLD=0.2
# Multi-scale DL fusion: off or a comma list, e.g. 0.5,1.0,1.5
# DETECTION_MULTISCALE=off
# Fusion strategy: smart_union | hysteresis | dl_only
# v3 ablation (runs/v3_app_ablation): dl_only @ 0.2 ≈ Test F1 0.58;
# smart_union/hysteresis collapsed F1 to ~0.04 — prefer dl_only with v3.
# DETECTION_FUSION=dl_only
# TTA: off | hflip | full | auto. Ablation: hflip ≈ same F1, ~2x slower on CPU.
# DETECTION_TTA=off
# Preprocessing toggles
# DETECTION_CLAHE=true
# DETECTION_HIST_MATCH=false
# DETECTION_SKIP_PREBLUR=         # auto: on in fullres_tiled mode
# Border pixels zeroed in mask cleanup (auto: 4 fullres / 12 downscaled)
# DETECTION_BORDER_MARGIN=
# Tiles per GPU forward pass (1 = no batching)
# DETECTION_TILE_BATCH=1
# Save a downsampled probability-map PNG per run for debugging
# DETECTION_SAVE_PROB_MAP=false
# Skip the deep model entirely on tiles that are fully unchanged between before/
# after (worst-cell LAB distance below this value — biggest single cell diff,
# not an average, so a small real change is never diluted away and missed).
# Big speed win on large images where most of the scene is static background;
# 0/unset = disabled (default). Start conservative (e.g. 3-5).
# DETECTION_SKIP_UNCHANGED_THRESHOLD=0

# --- smart_union fusion calibration (Delhi accuracy sprint, Day 4-6) ---
# Base classical-score percentile floor for smart_union fusion. Tuned from
# 0.92 -> 0.90 against real Delhi Sentinel-2 imagery: +53% mean F1 on the
# 16-pair labeled eval set, verified to pass the full synthetic regression
# suite (incl. the mandatory car-FP gate) unchanged. Lower values (0.75-0.85)
# scored even higher on Delhi but were REJECTED — they cause false positives
# on the parked_cars synthetic case. See runs/calibration/best_params.json
# for the full decision record (winning + rejected configs, with reasons).
# DETECTION_CL_Q_BASE=0.90
# Base DL confidence floor for smart_union fusion (default 0.36). Tested
# down to 0.10 against Delhi imagery — no meaningful effect at full-set scale
# (the pretrained AdaptFormer model's near-zero output on this coarse-GSD,
# out-of-domain imagery isn't a thresholding problem; needs fine-tuning, not
# calibration). Left at default. See runs/calibration/best_params.json.
# DETECTION_DL_FLOOR_BASE=0.15

# --- Research-integration channels (KPCAMNet / IR-MAD / SRCDNet / BIT_CD) ---
# KPCA deep features (KPCAMNet): auto (default; drives Feature-Based method),
# on (also adds a channel to the classical score map), off
# DETECTION_KPCA=auto
# Analysis-resolution cap for KPCA feature extraction (CPU bound)
# DETECTION_KPCA_MAX_SIDE=768
# IR-MAD change statistic: auto (default) | on (adds classical channel) | off
# DETECTION_IRMAD=auto
# Regression radiometric normalization from IR-MAD no-change pixels
# (off by default: reduced F1 on the LEVIR-CD gate; useful for strong
# sensor/illumination mismatch between acquisitions)
# DETECTION_IRMAD_NORM=false
# Resample GeoTIFF pairs to a common (coarser) GSD before detection
# DETECTION_GSD_HARMONIZE=true
# Relative GSD difference that triggers harmonization/preflight warning
# DETECTION_GSD_TOLERANCE=0.15
# Preflight overlap gate: below this fraction of the after-image footprint,
# a pair is rejected as not suitable for change detection (hard fail)
# DETECTION_MIN_OVERLAP_HARD=0.02
# Below this fraction, overlap is weak but detection still runs (soft warning)
# DETECTION_MIN_OVERLAP_WARN=0.20
# Ensemble AdaptFormer with BIT_CD (requires weights, see below; default off)
# DETECTION_ENSEMBLE=off
# BIT_CD LEVIR-CD checkpoint path. Download best_ckpt.pt from the BIT_CD
# repo's release links (https://github.com/justchenhao/BIT_CD) and place it at
# app/cd_models/weights/bit_cd_levir.pth (raw training checkpoints with a
# model_G_state_dict key are accepted as-is).
# BIT_CD_WEIGHTS=app/cd_models/weights/bit_cd_levir.pth

# Public URL for report links in emails (default http://localhost:8000 locally)
# PUBLIC_BASE_URL=http://localhost:8000

# Email notifications (optional). Password reset still uses this backend.
# EMAIL_API_URL=https://emailservice.managemybusinessess.com/api/email/send
# Login SMS OTP: POST ToMobile/Message to SMS_API_URL (default: same host as
# EMAIL_API_URL with /api/sms/send). Or set SMS_PROVIDER=fast2sms + SMS_API_KEY.
# SMS_API_URL=
# SMS_PROVIDER=
# SMS_API_KEY=
# LOGIN_OTP_LOG=1
# SMTP_HOST=smtp.gmail.com
# SMTP_PORT=587
# SMTP_USER=
# SMTP_PASS=

# Hugging Face model cache directory
# HF_HOME=./.hf_cache

# Department export API (optional — FR-08)
# DEPT_API_URL=https://dept.example.gov/api/changes
# DEPT_API_KEY=your-api-key

# DDA admin account (optional — UAT / ops)
# DDA_ADMIN_EMAIL=admin@example.com
# DDA_ADMIN_PASSWORD=change-me
# DDA_TRAINING_EXPORT_KEY=secret-key-for-fp-export