================================================================================
SUPPLEMENTARY MATERIAL — PHASE I MASTER CANDIDATE CATALOG
New River Valley, Belize — Maya Pottery Sherd Detection Study
University of Cincinnati, Department of Geography & GIS
Author: Benjamin Britton
================================================================================

ASSOCIATED ARTICLE
------------------
Britton, B. (2026). A Hue-Weighted Two-Phase Workflow for Automated Detection of
Maya Pottery Red Sherds in UAV Imagery. Remote Sensing (MDPI).

CONTENTS OF THIS ARCHIVE
--------------------------
  geom4999_enriched.csv          177,148-row Phase I master candidate catalog
  run_geom_full_4999.py          Reproduction script that generated the catalog
  README_Phase1_Master_Catalog.txt   This file


WHAT THIS DATASET IS
---------------------
geom4999_enriched.csv is the output of the Two-Phase Workflow's Phase I stage:
a full geometry-only inference pass across all 100 UAV survey images using a
Cascade Mask R-CNN model with ViTDet backbone (checkpoint model_0004999.pth).

No chromatic filtering has been applied. This file represents every candidate
object the geometric detector found, with all hue and saturation metrics
pre-computed and attached. It is the reproducible asset from which the Phase II
chromatic filter is applied to yield the final 1,647 detections reported in the
article.

As noted in the article: any re-parameterization of the Phase II chromatic
thresholds can be applied to this 177,148-candidate set in approximately
1.2 seconds per iteration, enabling rapid sensitivity analysis without
repeating the computationally intensive inference stage (~14 GPU-hours).


PROVENANCE
----------
  Dataset:          100-image UAV survey, New River Valley, Belize
  Sensor:           DJI drone, DNG raw imagery, 100 m AGL
  Image resolution: 5280 x 3956 px
  Tiling:           512 px tiles, 384 px stride, 128 px overlap
  Model:            Cascade Mask R-CNN, ViTDet backbone
  Checkpoint:       model_0004999.pth
                    (C:/d2/Outputs/output5_pseudo_rgb_transfer/)
  Confidence threshold: 0.05 (permissive — all geometry candidates retained)
  No deduplication applied in Phase I (tile-overlap duplicates preserved;
    cluster density is a quality signal exploited in Phase II scoring)
  UTM correction:   +0.5 px half-pixel centre offset applied to all centroids
  CRS:              EPSG:32616 (UTM Zone 16N, WGS84)
  Run date:         2026-03-18
  Script version:   geom_full_4999_v1


COLUMN DEFINITIONS (36 columns)
---------------------------------
  EVT_Rank            Integer rank by Geometric_Score (1 = highest confidence)
  Object_ID           Unique detection identifier (Flight_ID + Tile + Sherd)
  Flight_ID           Source DNG image identifier (e.g. DJI_0016)
  Tile_Num            Tile index within the flight image
  Sherd_Num           Detection index within the tile
  Combined_Confidence Raw model confidence score (0–1)
  Gold_Score          Composite score: 0.70 × Color_Score_Raw + 0.30 × Geom_Norm
  Geometric_Score     Normalised geometric confidence (0–1)
  Geom_Norm           Geometry score normalised within flight
  Color_Score_Raw     Chromatic similarity score, centred on 32.0° hue
  Color_Score_Penalized  Color score after limestone penalty (hues 40–45°)
  Hue_Deg             Mean hue of detection mask (degrees, 0–360)
  Saturation          Mean saturation of detection mask (0–100 scale)
  Value               Mean brightness of detection mask (0–100 scale)
  BBox_X              Bounding box left edge (pixels, image coordinates)
  BBox_Y              Bounding box top edge (pixels, image coordinates)
  BBox_W              Bounding box width (pixels)
  BBox_H              Bounding box height (pixels)
  BBox_Area           Bounding box area (pixels²)
  Size_Category       Ordinal size class (Small / Medium / Large)
  Cluster_Size        Number of tile detections within dedup radius
  UTM_Easting         Detection centroid easting, EPSG:32616 (metres)
  UTM_Northing        Detection centroid northing, EPSG:32616 (metres)
  Latitude            WGS84 decimal degrees
  Longitude           WGS84 decimal degrees
  Notes               Processing flags or anomaly notes
  Chip_Filename       Filename of 128 px chip image (chips not included here;
                        chips for 1,647 final detections available separately)
  RLE_Mask            Run-length encoded binary segmentation mask (COCO format)
  CIRC_HUE            Circular mean hue (atan2 method, degrees)
  MED_HUE             Median hue of mask pixels (degrees)
  STD_HUE             Circular standard deviation of hue (degrees)
  MEAN_SAT            Mean saturation of mask pixels (0–1 normalised)
  MED_SAT             Median saturation of mask pixels (0–1 normalised)
  MEAN_VAL            Mean brightness of mask pixels (0–1 normalised)
  MASK_PX             Total pixel count of segmentation mask
  HUE_SRC             Source of hue measurement ('mask' or 'bbox')


REPRODUCTION INSTRUCTIONS
--------------------------
Requirements:
  Python 3.8+, PyTorch 1.9+ with CUDA, Detectron2 (ViTDet build),
  numpy, opencv-python, scikit-image, pyproj, rawpy

Steps:
  1. Place run_geom_full_4999.py in the same directory as
     run_evt_locked_mpr01.py (the core MPRDetector pipeline).
  2. Ensure model_0004999.pth is at:
       C:/d2/Outputs/output5_pseudo_rgb_transfer/model_0004999.pth
     (or update WEIGHTS_PATH in the CONFIG dict).
  3. Ensure pre-processed HSV arrays are at G:/100pix/hsv_arrays/
     (one *_hsv_raw.npy per flight image).
  4. Run:
       python run_geom_full_4999.py
  5. Output writes to G:/100pix_EVT_geom4999/geom4999_summary.csv (727,991
     rows, all tile votes). The enriched 177,148-row file is produced by the
     downstream deduplication and enrichment step in the Phase II pipeline.

Approximate runtime: 12–14 hours on a single NVIDIA GPU (RTX class).
Re-running Phase II chromatic filtering only (from this catalog): ~1.2 sec.


NOTES
------
- HSV arrays are stored normalised 0–1 (float32), NOT 0–65535.
- The 727,991-row summary retains all tile votes (no dedup); the 177,148-row
  enriched file has duplicate tile-overlap detections collapsed to unique
  object centroids at 0.01 m radius.
- This catalog preceded the HITL adjudication step. The adjudicated
  validation dataset (FP_detections.shp, 786 records) is a separate
  supplementary file.


LICENSE / CITATION
-------------------
These data are released as supplementary material to the above article.
Please cite the article if you use this dataset or the associated code.

================================================================================
End of README
================================================================================
