ChangeJEPA: Predict Before You Compare for Remote Sensing Change Detection
Organizations: Johns Hopkins University
Abstract
Remote sensing change detection is typically formulated as Siamese feature comparison, where two images are encoded independently and a decoder determines which discrepancies correspond to meaningful change. This is challenging because the encoders preserve both true scene changes and nuisance variations such as illumination, seasonal, radiometric, and geometric differences. We introduce ChangeJEPA, a bitemporal joint-embedding predictive architecture that reformulates change detection as latent post-state prediction. Given a pre-change image and partial post-change evidence, ChangeJEPA predicts the full post-change representation under two complementary conditions: a prior hypothesis derived from the pre-change scene and an updated hypothesis conditioned on observed post-change context. A joint-embedding predictive loss aligns both hypotheses with the target latent representation, encouraging semantic consistency while suppressing transient appearance variations. Change is then inferred from the disagreement among the pre-change representation, the observed post-change target, and the predicted latent states, capturing complementary evidence of change, prior surprise, contextual update, and prediction reliability. Across SYSU-CD, LEVIR-CD+, and CDD, ChangeJEPA consistently outperforms its decoder matched Siamese baseline by up to 1.51 IoU points and achieves the best IoU and F1-score on all three benchmarks, demonstrating the advantage of latent post-state prediction over direct feature comparison. Page: https://buddhi19.github.io/ChangeJEPA
Figures & tables
| Evidence Streams | Decoder | SYSU-CD | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Variant | Morph. | OA | ||||||||||
| Siamese baseline | 73.61 | 84.80 | 92.78 | 84.26 | 85.34 | |||||||
| w/o prior prediction | 74.19 | 85.18 | 93.14 | 86.77 | 83.65 | |||||||
| w/o updated prediction | 74.11 | 85.13 | 92.76 | 82.59 | 87.82 | |||||||
| w/o direct evidence | 73.76 | 84.90 | 92.86 | 84.67 | 85.14 | |||||||
| w/o skip contexts | 73.88 | 84.98 | 92.88 | 84.56 | 85.40 | |||||||
| Method | SYSU-CD | LEVIR-CD+ | CDD | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| IoU | F1 | OA | Pre | Rec | IoU | F1 | OA | Pre | Rec | IoU | F1 | OA | Pre | Rec | |
| FC-EF ( Caye Daudt et al. 2018 ) | 64.35 | 78.31 | 89.58 | 76.91 | 79.77 | 58.69 | 73.97 | 97.98 | 78.01 | 70.32 | 85.26 | 92.04 | 98.06 | 93.17 | 90.94 |
| FC-Siam-Diff ( Caye Daudt et al. 2018 ) | 65.05 | 78.83 | 90.33 | 81.53 | 76.30 | 60.15 | 75.12 | 98.09 | 79.83 | 70.93 | 82.97 | 90.69 | 97.77 | 93.41 | 88.12 |
| FC-Siam-Conc ( Caye Daudt et al. 2018 ) | 63.85 | 77.94 | 89.67 | 78.52 | 77.36 | 62.28 | 76.75 | 98.12 | 77.37 | 76.15 | 84.70 | 91.72 | 98.01 | 94.31 | 89.26 |
| IFNet ( Zhang et al. 2020 ) | 67.42 | 80.54 | 90.84 | 80.67 | 80.41 | 68.18 | 81.08 | 98.47 | 81.76 | 80.41 | 87.35 | 93.25 | 98.36 | 94.74 | 91.80 |
| SNUNet ( Fang et al. 2022 ) | 53.32 | 69.56 | 88.27 | 89.70 | 56.80 | 62.78 | 77.14 | 98.18 | 79.14 | 75.24 | 67.04 | 80.27 | 95.50 | 87.59 | 74.07 |
| Evidence Streams | Decoder | Training | SYSU | LEVIR | |||||||||
| Variant | Morph. | WU | |||||||||||
| Siamese baseline | 73.61 | 84.80 | 74.89 | 85.64 | |||||||||
| w/o prior prediction | 74.19 | 85.18 | 74.64 | 85.48 | |||||||||
| w/o updated prediction | 74.11 | 85.13 | 73.69 | 84.85 | |||||||||
| w/o direct evidence | 73.76 | 84.90 | 74.73 | 85.54 | |||||||||
| w/o skip contexts | 73.88 | 84.98 | 75.66 | 86.14 | |||||||||
| Dataset | Method | 2 px | 4 px | 8 px | |||
|---|---|---|---|---|---|---|---|
| LEVIR-CD+ | ChangeJEPA-Baseline | 70.00 | 82.35 | 61.65 | 76.27 | 47.30 | 64.22 |
| ChangeJEPA | 71.37 | 83.29 | 62.90 | 77.23 | 48.50 | 65.32 | |
| SYSU-CD | ChangeJEPA-Baseline | 72.02 | 83.73 | 71.63 | 83.47 | 69.51 | 82.01 |
| ChangeJEPA | 74.34 | 85.28 | 73.51 | 84.73 | 69.59 | 82.07 | |
| Method | SYSU-CD | LEVIR-CD+ | CDD | |||
|---|---|---|---|---|---|---|
| Single pass, no TTA | ||||||
| Best competitor † | 73.95 | 85.04 | 75.59 | 86.10 | 95.82 | 97.87 |
| ChangeJEPA-Baseline | 72.97 | 84.37 | 74.51 | 85.39 | 95.23 | 97.56 |
| ChangeJEPA | 74.66 | 85.49 | 75.70 | 86.17 | 96.41 | 98.17 |
| With TTA (Table 2) | ||||||
| SYSU-CD | LEVIR-CD+ | |
| Mask seed varied (10 seeds, 70% hidden) | ||
| Std. of IoU | 0.003 | 0.002 |
| Largest | 0.005 | 0.003 |
| Hidden ratio varied (seed fixed): vs. 70% | ||
| 50% hidden | 0.003 | 0.013 |
| 60% hidden | 0.004 | 0.006 |
| Hidden-token filling | ||
|---|---|---|
| Mask embedding same-index token (ours) | 0 | 0 |
| Mask embedding only | 0.54 | 0.36 |
| Dataset | Released sample size | # Train | # Test |
|---|---|---|---|
| LEVIR-CD+ | 10,192 | 5,568 | |
| SYSU-CD | 16,000 | 4,000 | |
| CDD | 12,998 | 3,000 |
| Method | GFLOPs | Latency (ms) | FPS |
|---|---|---|---|
| FC-EF | 6.19 | 3.06 | 326.4 |
| FC-Siam-Diff | 8.46 | 4.15 | 241.1 |
| FC-Siam-Conc | 9.66 | 4.16 | 240.5 |
| IFNet | 157.86 | 29.30 | 34.1 |
| SNUNet | 8.46 | 3.78 | 264.7 |
| BIT | 17.41 | 13.18 | 75.9 |
| Family | Component | Endpoint | Weight |
|---|---|---|---|
| Illumination | Exposure | EV | |
| Gamma | |||
| Contrast | |||
| Colour temperature | K | ||
| Directional illumination | EV | ||
| Cast shadow | attenuation |
| Method | 25% | 50% | 75% | |||
|---|---|---|---|---|---|---|
| ChangeMamba | 65.63 | 79.25 | 53.11 | 69.38 | 40.28 | 57.43 |
| CDMamba | 63.64 | 77.78 | 50.65 | 67.24 | 43.30 | 60.43 |
| ChangeRWKV | 73.98 | 85.05 | 65.85 | 79.41 | 48.83 | 65.62 |
| ChangeCLIP | 70.45 | 82.66 | 54.49 | 70.54 | 41.34 | 58.50 |
| ChangeDINO | 73.98 | 85.04 | 63.51 | 77.69 | 44.23 | 61.33 |
Remote Sensing Change Detection
Remote sensing change detection finds and interprets differences across Earth-observation images acquired at different times.
See all 100 papers