geoai/geoai-workbench.git

			@@ -33,6 +33,7 @@
			MIN_MAX_DIMENSION = 512
			MAX_MAX_DIMENSION = 4096
			PROCESSING_MODES = {"auto", "image", "geotiff"}
			GENERIC_DIFF_THRESHOLD = 0.45


			def _to_uint8(data: np.ndarray) -> np.ndarray:
			@@ -171,13 +172,59 @@


			def _write_overlay(path: Path, before: np.ndarray, after: np.ndarray, mask: np.ndarray) -> None:
			_write_evidence_overlay(path, before, after, mask, None)


			def _write_evidence_overlay(path: Path, before: np.ndarray, after: np.ndarray, model_mask: np.ndarray, generic_mask: np.ndarray \| None) -> None:
			left = before.copy()
			right = after.copy()
			if generic_mask is not None:
			cyan = np.zeros_like(right)
			cyan[..., 1] = 220
			cyan[..., 2] = 255
			right[generic_mask] = (right[generic_mask].astype(np.float32) * 0.45 + cyan[generic_mask].astype(np.float32) * 0.55).astype(np.uint8)
			red = np.zeros_like(right)
			red[..., 0] = 255
			right[mask] = (right[mask].astype(np.float32) * 0.45 + red[mask].astype(np.float32) * 0.55).astype(np.uint8)
			right[model_mask] = (right[model_mask].astype(np.float32) * 0.45 + red[model_mask].astype(np.float32) * 0.55).astype(np.uint8)
			separator = np.full((before.shape[0], 8, 3), 235, dtype=np.uint8)
			Image.fromarray(np.concatenate([left, separator, right], axis=1)).save(path, quality=92)


			def _compute_generic_difference_mask(before: np.ndarray, after: np.ndarray, valid: np.ndarray, *, threshold: float = GENERIC_DIFF_THRESHOLD) -> tuple[np.ndarray, dict[str, Any]]:
			"""Create a conservative visual-difference candidate layer.

			This is deliberately separate from ChangeStar's semantic mask: it highlights
			strong local RGB differences such as moved vehicles or people, while
			removing tiny noise, very large illumination regions, and invalid borders.
			"""
			if before.shape != after.shape or before.shape[:2] != valid.shape:
			raise ValueError("before, after, and valid must share the same spatial shape")
			difference = np.mean(np.abs(before.astype(np.float32) - after.astype(np.float32)), axis=2) / 255.0
			difference = cv2.GaussianBlur(difference, (3, 3), 0)
			candidate = ((difference >= threshold) & valid).astype(np.uint8)
			candidate = cv2.morphologyEx(candidate, cv2.MORPH_OPEN, np.ones((3, 3), dtype=np.uint8))
			candidate = cv2.morphologyEx(candidate, cv2.MORPH_CLOSE, np.ones((5, 5), dtype=np.uint8))
			count, labels, stats, _ = cv2.connectedComponentsWithStats(candidate, connectivity=8)
			min_area = max(32, int(candidate.size * 0.00003))
			max_area = max(min_area, int(candidate.size * 0.08))
			cleaned = np.zeros_like(candidate)
			kept = 0
			for component in range(1, count):
			x, y, width, height, area = (int(value) for value in stats[component])
			touches_edge = x <= 0 or y <= 0 or x + width >= candidate.shape[1] or y + height >= candidate.shape[0]
			if min_area <= area <= max_area and not touches_edge:
			cleaned[labels == component] = 1
			kept += 1
			details = {
			"threshold": threshold,
			"minimum_component_pixels": min_area,
			"maximum_component_pixels": max_area,
			"components_before_filter": max(0, count - 1),
			"components_after_filter": kept,
			"candidate_pixels": int(cleaned.sum()),
			"candidate_pixel_ratio": round(float(cleaned.mean()), 6),
			}
			return cleaned > 0, details


			def _feature_summary(mask: np.ndarray, probability: np.ndarray, transform: Affine, georeferenced: bool) -> list[dict[str, Any]]:
			@@ -311,6 +358,8 @@
			output_crs = before_info["crs"] if actual_mode == "geotiff" else None
			_write_rgb_geotiff(work_dir / "before.tif", before_small, transform, output_crs)
			_write_rgb_geotiff(work_dir / "after_registered.tif", after_small, transform, output_crs)
			Image.fromarray(before_small).save(output_dir / "before_processed_preview.jpg", quality=92)
			Image.fromarray(after_small).save(output_dir / "after_registered_preview.jpg", quality=92)

			from geoai import ChangeStarDetection

			@@ -325,10 +374,13 @@
			for component in range(1, count):
			if int(stats[component, cv2.CC_STAT_AREA]) < min_area:
			cleaned[labels == component] = 0
			generic_mask, generic_details = _compute_generic_difference_mask(before_small, after_small, valid_small)
			_write_raster(output_dir / "change_probability.tif", probability, transform, "float32", output_crs)
			_write_raster(output_dir / "change_mask_raw.tif", raw_mask, transform, "uint8", output_crs)
			_write_raster(output_dir / "change_mask.tif", cleaned, transform, "uint8", output_crs)
			_write_overlay(output_dir / "change_overlay.jpg", before_small, after_small, cleaned > 0)
			_write_raster(output_dir / "generic_difference_mask.tif", generic_mask.astype(np.uint8) * 255, transform, "uint8", output_crs)
			_write_evidence_overlay(output_dir / "change_overlay.jpg", before_small, after_small, cleaned > 0, generic_mask)
			_write_overlay(output_dir / "change_model_overlay.jpg", before_small, after_small, cleaned > 0)
			vector_path = output_dir / "changes.geojson"
			features, vector_count = vectorize_cleaned_mask(output_dir / "change_mask.tif", output_dir / "change_probability.tif", vector_path)
			from rectangularize_vectors import rectangularize_vector
			@@ -371,6 +423,7 @@
			"raw_changed_pixels": int((raw_mask > 0).sum()),
			"changed_pixels": int((cleaned > 0).sum()),
			"changed_pixel_ratio": round(float((cleaned > 0).mean()), 6),
			"generic_difference": generic_details,
			"vector_feature_count": vector_count,
			"rectangle_feature_count": rectangle_count,
			"georeferenced": actual_mode == "geotiff",
			@@ -383,7 +436,11 @@
			"probability_raster": "change_probability.tif",
			"raw_mask_raster": "change_mask_raw.tif",
			"mask_raster": "change_mask.tif",
			"generic_difference_mask": "generic_difference_mask.tif",
			"overlay": "change_overlay.jpg",
			"model_overlay": "change_model_overlay.jpg",
			"before_processed_preview": "before_processed_preview.jpg",
			"after_registered_preview": "after_registered_preview.jpg",
			"vector": "changes.geojson",
			"rectangle_vector": "changes_rectangles.geojson",
			"rectangle_vector_wgs84": "changes_rectangles_wgs84.geojson" if actual_mode == "geotiff" else None,