Reject blank positive imagery in training loop

scripts/audit_operator_yolo_dataset_quality.py

@@ -4,7 +4,7 @@ from __future__ import annotations
 import argparse
 import json
 import statistics
-from collections import Counter, defaultdict
+from collections import Counter
 from pathlib import Path
 from typing import Any
 
@@ -246,6 +246,7 @@ def build_audit(summary: dict[str, Any], summary_path: Path, args: argparse.Name
     ]
     positive_tiles = [tile for tile in tiles if tile not in negative_tiles]
     train_negative_tiles = [tile for tile in negative_tiles if tile.get("split") == "train"]
+    low_variance_positive_tiles = [tile for tile in positive_tiles if tile.get("low_visual_variance")]
     val_positive_samples = {
         str(tile.get("sample_slug"))
         for tile in positive_tiles
@@ -288,6 +289,15 @@ def build_audit(summary: dict[str, Any], summary_path: Path, args: argparse.Name
                 f"gate is {args.max_repeated_negative_share:.3f}."
             ),
         )
+    if low_variance_positive_tiles:
+        add_warning(
+            warnings,
+            "positive_tiles_have_low_visual_variance",
+            (
+                f"{len(low_variance_positive_tiles)} positive tiles are visually blank/low-variance; "
+                "the imagery source does not support their labels."
+            ),
+        )
     if label_stats["missing_label_file_count"]:
         add_warning(
             warnings,
@@ -337,6 +347,7 @@ def build_audit(summary: dict[str, Any], summary_path: Path, args: argparse.Name
         "train_tile_count": split_counts.get("train", 0),
         "val_tile_count": split_counts.get("val", 0),
         "train_negative_tile_count": len(train_negative_tiles),
+        "low_variance_positive_tile_count": len(low_variance_positive_tiles),
         "repeated_background_negative_tile_count": repeated_negative_count,
         "repeated_background_negative_share_of_negatives": repeated_negative_share,
         "positive_tile_share": len(positive_tiles) / len(tiles) if tiles else 0.0,
@@ -371,6 +382,10 @@ def build_recommendations(warnings: list[dict[str, str]]) -> list[str]:
         )
     if "label_files_missing" in codes or "invalid_label_rows" in codes:
         recommendations.append("Regenerate the YOLO tile dataset and review exporter path/label integrity.")
+    if "positive_tiles_have_low_visual_variance" in codes:
+        recommendations.append(
+            "Reject the raster product for affected AOIs or replace it with an officially complete imagery edition before training."
+        )
     if not recommendations:
         recommendations.append("Dataset audit passed the configured gates; continue with benchmarked training.")
     return recommendations
@@ -389,6 +404,7 @@ def write_markdown(report: dict[str, Any], path: Path) -> None:
         f"- Samples: {report['sample_count']} ({report['positive_sample_count']} positive, {report['background_sample_count']} background)",
         f"- Repeated background negative share: {report['repeated_background_negative_share_of_negatives']:.3f}",
         f"- Minimum visible label ratio: {format_optional_float(report.get('min_label_visible_ratio'))}",
+        f"- Blank/low-variance positive tiles: {report['low_variance_positive_tile_count']}",
         "",
         "## Label Quality",
         "",