diff --git a/tools/extract/leonida/classify.py b/tools/extract/leonida/classify.py
new file mode 100644
index 0000000..af1e12f
--- /dev/null
+++ b/tools/extract/leonida/classify.py
@@ -0,0 +1,43 @@
+import numpy as np
+from scipy import ndimage as ndi
+from skimage.color import rgb2lab
+
+PALETTE = {
+    "wasser":     (44, 103, 164),
+    "land":       (192, 216, 128),
+    "gruen":      (139, 195, 74),
+    "sand":       (232, 216, 136),
+    "stadt":      (216, 216, 216),
+    "gebaeude":   (176, 176, 176),
+    "strasse":    (80, 80, 80),
+    "weg":        (192, 152, 96),
+    "spekulativ": (198, 96, 96),
+}
+CLASSES = {name: i for i, name in enumerate(PALETTE)}
+UNKNOWN = 255
+MAX_DELTA_E = 14.0     # darüber gilt ein Pixel als unbekannt (Beschriftung, Icon, Kante)
+
+
+def classify(rgb: np.ndarray, report_unknown: bool = False):
+    """Ordnet jedes Pixel der nächsten Palettenfarbe zu; Ausreißer werden aufgefüllt.
+
+    Mit report_unknown=True kommt zusätzlich der Anteil der Pixel zurück, die vor dem
+    Auffüllen zu keiner Legendenfarbe passten. Aufgabe 6 bricht darüber bei über 5 % ab.
+    """
+    lab = rgb2lab(rgb.astype(np.float32) / 255.0)
+    ref = rgb2lab(np.array([list(PALETTE.values())], dtype=np.float32) / 255.0)[0]
+    dist = np.linalg.norm(lab[:, :, None, :] - ref[None, None, :, :], axis=3)
+    labels = dist.argmin(axis=2).astype(np.uint8)
+    unknown = dist.min(axis=2) > MAX_DELTA_E
+    share = float(unknown.mean())
+    if unknown.any():
+        # Nächstes bekanntes Pixel übernehmen.
+        idx = ndi.distance_transform_edt(unknown, return_distances=False, return_indices=True)
+        labels = labels[tuple(idx)]
+    return (labels, share) if report_unknown else labels
+
+
+def class_shares(labels: np.ndarray) -> dict[str, float]:
+    total = labels.size
+    counts = np.bincount(labels.ravel(), minlength=len(PALETTE))
+    return {name: float(counts[i]) / total for name, i in CLASSES.items()}
diff --git a/tools/extract/tests/test_classify.py b/tools/extract/tests/test_classify.py
new file mode 100644
index 0000000..b562ec0
--- /dev/null
+++ b/tools/extract/tests/test_classify.py
@@ -0,0 +1,36 @@
+import numpy as np
+from leonida.classify import CLASSES, PALETTE, class_shares, classify
+
+def block(color, size=8):
+    return np.full((size, size, 3), color, dtype=np.uint8)
+
+def test_pure_colors_hit_their_class():
+    for name, color in PALETTE.items():
+        labels = classify(block(color))
+        assert (labels == CLASSES[name]).all(), name
+
+def test_jpeg_noise_still_classifies():
+    rng = np.random.default_rng(0)
+    noisy = block(PALETTE["strasse"], 32).astype(np.int16)
+    noisy += rng.integers(-10, 11, noisy.shape)          # JPEG-Rauschen
+    labels = classify(np.clip(noisy, 0, 255).astype(np.uint8))
+    assert (labels == CLASSES["strasse"]).mean() > 0.98
+
+def test_unknown_pixels_are_filled_from_neighbours():
+    img = block(PALETTE["stadt"], 16)
+    img[7:9, 7:9] = (255, 0, 255)                         # Beschriftung, in keiner Klasse
+    labels = classify(img)
+    assert (labels == CLASSES["stadt"]).all()
+
+def test_reports_the_share_of_unknown_pixels():
+    img = block(PALETTE["stadt"], 16)
+    img[7:9, 7:9] = (255, 0, 255)              # 4 von 256 Pixeln passen zu keiner Farbe
+    labels, share = classify(img, report_unknown=True)
+    assert (labels == CLASSES["stadt"]).all()
+    assert abs(share - 4 / 256) < 1e-9
+
+def test_shares_sum_to_one():
+    img = np.concatenate([block(PALETTE["wasser"], 8), block(PALETTE["land"], 8)], axis=1)
+    shares = class_shares(classify(img))
+    assert abs(sum(shares.values()) - 1.0) < 1e-9
+    assert abs(shares["wasser"] - 0.5) < 1e-9