You can deploy this model and start earning money today!

Rohit7901/
Cross-model-change-detection

Name: Cross-model-change-detection
Availability: InStock

Views

No views yet

EO–SAR Cross-Modal Change Detection

Pixel-level binary change detection between pre-event optical (EO) and post-event radar (SAR) satellite images.
Outputs a mask where 1 = change (damage/flood) and 0 = no change.

Model

V9 (primary): Dual-encoder ResNet34-UNet with CBAM

EO encoder: ResNet34 (ImageNet-pretrained), 3-channel RGB input
SAR encoder: ResNet34 (ImageNet-pretrained, conv1 adapted), 6 engineered feature channels
Fusion: Cross-modal feature fusion at bottleneck + multi-scale skip connections with CBAM at bottleneck and high-level decoder stages (d3, d2)
Loss: Hybrid Focal BCE + Focal Tversky (α=0.2, β=0.8, γ=1.5) tuned for recall on the change class
SAR preprocessing: Lee adaptive speckle filter applied post-normalisation

All implemented in pure PyTorch (no segmentation_models_pytorch).

Usage

python

import torch
from huggingface_hub import hf_hub_download
from model import DualEncoderUNet

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Download V9 weights
ckpt_path = hf_hub_download(
    repo_id="Rohit7901/galaxeye-change-detection",
    filename="best_v9.pth",
)

# Instantiate model
model = DualEncoderUNet(
    eo_channels=3,
    sar_channels=6,
    num_classes=1,
).to(device)

model.load_state_dict(torch.load(ckpt_path, map_location=device, weights_only=True))
model.eval()

# 4-pass TTA inference (images: B, C, H, W tensor on device)
with torch.no_grad():
    pred_orig   = torch.sigmoid(model(images))

    pred_hflip  = torch.sigmoid(model(torch.flip(images, dims=[3])))
    pred_hflip  = torch.flip(pred_hflip, dims=[3])

    pred_vflip  = torch.sigmoid(model(torch.flip(images, dims=[2])))
    pred_vflip  = torch.flip(pred_vflip, dims=[2])

    pred_hvflip = torch.sigmoid(model(torch.flip(images, dims=[2, 3])))
    pred_hvflip = torch.flip(pred_hvflip, dims=[2, 3])

    avg_pred    = (pred_orig + pred_hflip + pred_vflip + pred_hvflip) / 4.0
    binary_mask = (avg_pred > 0.45).float()

Weights

Primary

File	Val IoU	Test IoU	Description
`best_v9.pth`	0.6050	0.3230	PRIMARY — dual-encoder ResNet34×2, 9-ch, 1024px, CBAM

Single-encoder baselines

Earlier experiments with a single ResNet34-UNet encoder, kept for comparison.

File	Val IoU	Test IoU	Description
`best_v5.pth`	0.4906	0.0080	9-ch, γ corrected to 1.33, 1024px
`best_v4.pth`	0.3137	0.0348	9-ch + NGRDI (normalisation bug present)
`best_v3.pth`	0.4198	0.0180	7-ch, 512px crop ablation

Loading a single-encoder checkpoint:

python

from model import ResNet34UNet

model = ResNet34UNet(in_channels=9, num_classes=1)
model.load_state_dict(
    torch.load("best_v5.pth", map_location="cpu", weights_only=True)
)
model.eval()

Results

V9 — Primary (dual encoder + CBAM)

Split	IoU	Precision	Recall	F1
Val (scenes 01–08)	0.6050	0.6763	0.8517	0.7539
Test (scene 09, unseen geography)	0.3230	0.4197	0.5837	0.4883

V5 — Single-encoder baseline

Split	IoU	Precision	Recall	F1
Val (scenes 01–08)	0.4906	0.6656	0.6511	0.6583
Test (scene 09, unseen geography)	0.0080	0.0276	0.0111	0.0159

V9 vs V5 on the unseen test scene:

Metric	V5	V9	Absolute gain
IoU	0.0080	0.3230	+0.315 (~40×)
Precision	0.0276	0.4197	+0.392
Recall	0.0111	0.5837	+0.573
F1	0.0159	0.4883	+0.472

The V5→V9 test IoU improvement comes from three architectural changes applied incrementally: switching to a symmetric dual-encoder to handle EO/SAR distribution mismatch, adding CBAM at the bottleneck and decoder stages for false-positive suppression, and applying Lee adaptive speckle filtering to improve SAR feature quality.

The remaining val→test gap (~0.28 IoU) is a geographic generalisation problem. Scene 09 is an entirely different region from the 8 training scenes. Threshold sweep analysis confirmed ~96% of the gap is generalisation failure, not calibration error.

License

Apache 2.0 — see the repository LICENSE file.
Please cite the associated README if you use this model in academic work.

Rohit7901/Cross-model-change-detection