Hugging Face's logo

Spaces:
GeospatiaProject
/
geospatial1
Sleeping

App Files Community
geospatial1
File size: 4,178 Bytes
d1ac326
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
interface Row {
  model: string;
  params: string;
  f1: string;
  iou: string;
  ap: string;
  note?: string;
}

// Real LEVIR-CD test results (threshold selected on val, applied to test) — from the committed
// comparison in docs/results/. This is the genuine project story; the served ONNX may be a
// placeholder until the trained bundles are staged.
const ROWS: Row[] = [
  { model: "FC-Siam-diff (baseline)", params: "0.83M", f1: "0.886", iou: "0.796", ap: "0.932" },
  {
    model: "Siamese-SegFormer / MiT-b2 (diff)",
    params: "24.72M",
    f1: "0.911",
    iou: "0.836",
    ap: "0.943",
    note: "ImageNet-pretrained strong model",
  },
  {
    model: "DINOv2-base frozen linear-probe",
    params: "1.64M",
    f1: "0.889",
    iou: "0.800",
    ap: "0.924",
    note: "frozen FM features only",
  },
  {
    model: "DINOv2-base + LoRA",
    params: "2.82M",
    f1: "0.913",
    iou: "0.839",
    ap: "0.946",
    note: "foundation-model tier — headline",
  },
];

export default function ModelCard() {
  return (
    <div className="card-page">
      <div className="card-inner">
        <p className="card-eyebrow">Model card · Track A · high-res aerial</p>
        <h2>Siamese change detection on LEVIR-CD</h2>
        <p className="card-lead">
          Weight-shared Siamese change-detection models on 0.5&nbsp;m RGB aerial imagery. Two dates of
          the same place go in; a per-pixel building-change map comes out. Three tiers are compared on
          the identical LEVIR-CD test split through one evaluation harness.
        </p>

        <table className="metrics-table">
          <thead>
            <tr>
              <th>Model</th>
              <th>Trainable</th>
              <th>F1</th>
              <th>IoU</th>
              <th>AP</th>
            </tr>
          </thead>
          <tbody>
            {ROWS.map((r) => (
              <tr key={r.model} className={r.note?.includes("headline") ? "row-headline" : ""}>
                <td>
                  {r.model}
                  {r.note && <span className="row-note">{r.note}</span>}
                </td>
                <td className="num">{r.params}</td>
                <td className="num">{r.f1}</td>
                <td className="num">{r.iou}</td>
                <td className="num">{r.ap}</td>
              </tr>
            ))}
          </tbody>
        </table>

        <h3>The defensible claim — parameter efficiency</h3>
        <p className="muted">
          DINOv2-base + LoRA reaches F1&nbsp;0.913 with only <strong>2.82M</strong> trainable
          parameters versus the SegFormer strong model&apos;s <strong>24.72M</strong> at F1&nbsp;0.911
          — a statistical tie on accuracy (within noise) at ~9× fewer trainable params. The frozen
          linear-probe (1.64M, decoder only) already ≈ the baseline, so the self-supervised
          representation carries most of the change signal and LoRA supplies the adaptation lift.
        </p>

        <h3>Honest limitations</h3>
        <ul className="muted">
          <li>
            Per-scene F1 has high variance (mean ≈ 0.77, std ≈ 0.31, min 0.00). The foundation model
            lifts the mean but does <em>not</em> fix the hardest small/subtle-change tiles.
          </li>
          <li>
            Overall pixel accuracy is deliberately not reported: change is a tiny pixel fraction, so
            &quot;predict no change&quot; scores ~99% and is meaningless. Metrics are change-class only.
          </li>
          <li>Trained weights inherit LEVIR-CD research/non-commercial terms — showcase use only.</li>
        </ul>

        <h3>Domain gap — why the live mode is a different model</h3>
        <div className="callout">
          <p className="muted">
            These models are trained on 0.5&nbsp;m aerial imagery and do <strong>not</strong> transfer
            to 10&nbsp;m Sentinel-2. The live-AOI mode (a later milestone) uses a Sentinel-2-native
            model so it produces meaningful output on satellite scenes — the domain split is a design
            decision, not an afterthought.
          </p>
        </div>
      </div>
    </div>
  );
}