ChangeJEPA: Predict Before You Compare for Remote Sensing Change Detection

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.

Key idea: a difference becomes more informative when it can be related to an expectation. ChangeJEPA estimates the post-change scene from the earlier observation, updates that estimate using visible post-change context, and examines how both predictions relate to the observed scenes. Five complementary comparisons reveal change that is represented in the prediction, introduced through contextual revision, or still unexpected.

Predict Before You Compare

A pixel-wise or feature-wise difference between two acquisitions answers one question: where do the images differ? Images acquired at different times can differ even when the underlying scene has not meaningfully changed, so a difference alone is an ambiguous signal. ChangeJEPA relates each difference to an expectation, organised around three questions answered in a common post-change latent space.

1 · Expectation

What should the scene look like now?

The prior prediction q̂₁ is a conditional estimate of the complete post-change representation, formed from the pre-change image alone.

2 · Revision

How does new context revise it?

The updated prediction q̂₂ revises that estimate using the visible subset of post-change patches.

3 · Surprise

What remains unexpected?

Comparing the observed state q₂ with both predictions exposes departures at two levels: before and after contextual updating.

What remains unexpected is an important signal of change. A newly observed scene may depart from the historical estimate, and some discrepancies may persist even after the visible post-change evidence has been incorporated. Genuine change can, however, also become explainable: if the visible patches reveal a new building, the updated predictor can incorporate it, and its residual against the observed state may then be small. The building is still evident in the difference between the updated prediction and the pre-change scene, and in the revision of the prior.

Successful prediction does not make a genuine change disappear; it changes where the evidence appears. The five evidence streams are designed to capture it wherever it appears.

Why Prediction Helps

The dominant paradigm in remote sensing change detection encodes two observations independently, compares their representations, and lets a decoder infer change from the resulting bitemporal discrepancy. These discrepancies are inherently ambiguous: illumination, shadows, seasonal variation, viewing geometry, radiometric differences, and registration errors can produce responses similar to genuine semantic change, so the decoder must separate true changes from pseudo-changes after they have already become entangled in the feature representations.

ChangeJEPA gives the decoder additional reference points. Building on joint-embedding predictive architectures, which learn semantic structure by predicting latent representations rather than pixels, it learns a prior post-change estimate from the pre-change image and an updated estimate conditioned on partial post-change evidence. Both predictions are trained against the same complete post-change target, which encourages them to express their estimates in a common post-change latent space. The complete post-change image remains available to the detector through this target; partial visibility applies only to the updated prediction branch.

This predictive reference gives temporal differences a richer interpretation: how the observed scene departs from an initial expectation, how new evidence revises that expectation, and what remains inconsistent after updating. The decoder combines these relationships with direct bitemporal evidence when deciding whether a difference indicates meaningful change.

The intended benefit is more selective change detection. Rather than treating a large discrepancy as sufficient evidence, the model learns to interpret discrepancies through their relations with observed and predicted states. Relative to the decoder-matched Siamese baseline, ChangeJEPA achieves higher precision on SYSU-CD and LEVIR-CD+, consistent with fewer false detections. The ablations on SYSU-CD indicate where this comes from: removing the updated-prediction relations lowers precision (86.37 → 82.59), whereas removing the prior-prediction relations lowers recall (85.22 → 83.65).

Siamese comparison

Encodes I₁ and I₂ independently and decodes their difference. The discrepancy shows where the images differ, but not whether the difference is a meaningful scene change.

Latent post-state prediction

Predicts the post-change state, first from I₁ alone and then with partial I₂, and compares observation with prediction in a shared, semantically aligned latent space.

ChangeJEPA Architecture

**Figure 2:** Overall architecture of ChangeJEPA. BiT-JEPA employs two weight-sharing context branches f_θ and a target encoder f_θ̄ to represent the complete pre-change image, the visible post-change patches, and the complete post-change image. Bitemporal context fusion forms a prior context z₁ with 0% post-change visibility and an updated context z₂ with p% visible post-change evidence. Weight-sharing predictors g_φ map these contexts to complete post-change predictions q̂₁ and q̂₂, both aligned with the target q₂ through L_JEPA. The Change Evidence Module constructs five complementary comparisons, and a four-stage MaxViT decoder with gated multi-level bitemporal skip contexts and differentiable morphological refinement produces the change map. *Click to enlarge.*

Bitemporal JEPA

A Vision Transformer target encoder, an exponential moving average of the context encoder, processes the complete post-change image and provides the latent prediction target q₂. The shared context encoder processes the complete pre-change image and a visible subset of post-change patches, obtained by masking large rectangular blocks of the post-change patch grid with a hidden ratio sampled from [0.6, 0.8]. Hidden post-change locations are filled with a learnable mask embedding combined with the pre-change token at the same grid location. A MaxViT fusion block forms a prior context (no post-change patches visible) and an updated context (visible post-change patches), and a shared predictor maps both to complete post-change predictions. The predictive loss aligns both predictions with the detached, layer-normalised target: L_JEPA = ℓ(q̂₂, q̄₂) + λ_prior ℓ(q̂₁, q̄₂), with q̄₂ = LN(sg(q₂)).

Change Evidence Module

Rather than relying on a single bitemporal difference, the module constructs five evidence streams. For two feature maps a and b, the pairwise descriptor concatenates both features, their absolute difference, and their cosine-distance map.

Stream	Relation	Evidence	Question it helps the decoder answer
e₁	x₁, q̂₂	Predicted change	What change is represented in the updated estimate?
e₂	x₁, q₂	Direct observed change	What differences are directly observed?
e₃	q̂₁, q̂₂	Change discrepancy	How does new context revise the initial estimate?
e₄	q̂₁, q₂	Prior surprise	What departs from the historical expectation?
e₅	q̂₂, q₂	Update reliability	What remains unexplained after updating?

The streams offer complementary views of the same scene transition. ChangeJEPA learns to combine them, so that both contextually explained change and unresolved surprise contribute to the final change map.

Change decoder and decoder-matched baseline

The aggregated evidence is decoded by four MaxViT stages with PixelShuffle upsampling. Multi-level bitemporal skip contexts from the eighth, sixteenth, and twenty-fourth transformer blocks are injected into the second to fourth stages through learnable gates, and a differentiable morphological refiner (soft opening and closing followed by a residual refiner) produces the final change map. To assess the predictive formulation, a decoder-matched Siamese baseline retains the same I-JEPA initialisation, evidence aggregation, decoder, skip contexts, and refiner, but replaces the five predictive streams with a single Siamese evidence stream.

Architecture of the decoder-matched ChangeJEPA baseline. — **Figure 4:** Architecture of the decoder-matched ChangeJEPA baseline, which replaces latent post-state prediction with direct Siamese comparison of the two encoded images.

Quantitative Results

ChangeJEPA is evaluated on three public binary change-detection benchmarks: SYSU-CD (heterogeneous urban and land-cover transitions), LEVIR-CD+ (very-high-resolution building changes), and CDD (pronounced seasonal appearance variation). All methods are trained and evaluated on the same train–test split. ChangeJEPA achieves the highest IoU, F1-score, and OA on all three benchmarks. Relative to the decoder-matched baseline, the improvement on SYSU-CD and LEVIR-CD+ is driven mainly by higher precision (86.37 vs 84.26 and 85.90 vs 83.90) at a small cost in recall (85.22 vs 85.34 and 86.89 vs 87.47), consistent with fewer false detections; on CDD, both precision and recall improve despite substantial seasonal variation.

Method	SYSU-CD					LEVIR-CD+					CDD
Method	IoU	F1	OA	Pre	Rec	IoU	F1	OA	Pre	Rec	IoU	F1	OA	Pre	Rec
FC-EF	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	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	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	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	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
BIT	64.79	78.63	89.82	77.82	79.46	66.43	79.83	98.31	77.89	81.87	87.91	93.57	98.44	95.10	92.08
ChangeFormer	66.03	79.54	90.60	81.74	77.46	64.74	78.60	98.25	78.32	78.87	94.34	97.09	99.28	97.48	96.70
ChangeMamba	71.67	83.50	92.63	88.45	79.07	69.63	82.09	98.50	80.17	84.12	95.82	97.87	99.47	97.85	97.88
CDMamba	68.27	81.14	91.38	83.79	78.65	67.65	80.71	98.43	80.86	80.56	89.82	94.64	98.68	94.90	94.37
ChangeRWKV	69.61	82.08	92.02	87.30	77.45	74.96	85.69	98.84	86.50	84.90	92.80	96.27	99.08	96.52	96.02
ChangeCLIP	72.40	83.99	92.64	86.18	81.91	73.60	84.79	98.76	84.73	84.86	95.64	97.77	99.45	97.86	97.69
ChangeDINO	73.95	85.04	92.61	81.35	89.07	75.59	86.10	98.85	84.42	87.41	92.59	96.15	99.05	96.41	95.90
ChangeJEPA-Baseline	73.61	84.80	92.78	84.26	85.34	74.89	85.64	98.81	83.90	87.47	95.48	97.69	99.43	97.07	98.32
ChangeJEPA	75.12	85.79	93.34	86.37	85.22	76.05	86.39	98.88	85.90	86.89	96.56	98.25	99.57	97.76	98.75

Table 1: Quantitative comparison on SYSU-CD, LEVIR-CD+, and CDD (%). Best in bold green, second best underlined.

Benchmark

Metric

Figure 5: Comparison of all methods in Table 1 on the selected benchmark and metric.

Ablation Study

The prior and updated predictions contribute complementary information. Removing the updated-prediction relations shifts the model towards higher recall but lower precision (87.82 / 82.59), producing broader and less selective detections; removing the prior-prediction relations produces the opposite tendency (83.65 / 86.77), missing changes that depart from the historical expectation. Direct bitemporal evidence remains essential and causes the strongest degradation among the evidence ablations when removed: predictive modelling complements rather than replaces direct comparison. The multi-level skip contexts improve spatial localisation, while the morphological refiner provides a smaller but consistent improvement.

Variant	Evidence streams					Decoder		SYSU-CD
Variant	e₁	e₂	e₃	e₄	e₅	Skips	Morph.	IoU	F1	OA	Pre	Rec
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
w/o morphological refiner	✓	✓	✓	✓	✓	✓		74.97	85.69	93.22	85.33	86.06
ChangeJEPA	✓	✓	✓	✓	✓	✓	✓	75.12	85.79	93.34	86.37	85.22

Table 2: Ablation study on SYSU-CD (%). The Siamese baseline uses a single direct comparison of the two encoded images in place of the predictive evidence streams; an empty entry indicates that the corresponding component is omitted.

Qualitative Results

Change maps on test pairs from LEVIR-CD+ (full 1024×1024 scenes, reassembled from the 256×256 test crops) and SYSU-CD, compared with the decoder-matched baseline and with ChangeMamba, ChangeCLIP, and ChangeDINO. White, red, blue, and black denote true positives, false negatives, false positives, and true negatives, respectively.

LEVIR-CD+ · scene 881

LEVIR-CD+ · scene 947

SYSU-CD · 01030

SYSU-CD · 00316

SYSU-CD · 01081

SYSU-CD · 01818

Figure 6: Qualitative comparison on test pairs from LEVIR-CD+ and SYSU-CD. ChangeJEPA uses the inference protocol of the paper; the compared methods use their released inference settings.

**Figure 7:** Qualitative comparison on LEVIR-CD+, SYSU-CD, and CDD. White, red, and blue denote true positives, false negatives, and false positives. On LEVIR-CD+, ChangeJEPA recovers compact and disconnected building changes while suppressing responses from neighbouring roofs and roads; on SYSU-CD, it better preserves narrow and concave unchanged regions within large change areas; on CDD, it retains thin structures and irregular contours despite strong seasonal and radiometric differences.

**Figure 8:** Additional qualitative comparison on CDD, LEVIR-CD+, and SYSU-CD. From left to right: pre-change image, post-change image, ChangeMamba, ChangeCLIP, ChangeDINO, the decoder-matched baseline, ChangeJEPA, and the ground-truth mask. White, black, blue, and red denote true positives, true negatives, false positives, and false negatives.

Latent Predictions and Change Evidence

The prior prediction q̂₁ represents the post-state inferred from I₁ alone, whereas the updated prediction q̂₂ revises this expectation using visible post-change context. The corresponding cosine-distance maps exhibit distinct spatial responses: the observed bitemporal relation identifies where the scene differs, the prior–target relation expresses departure from the expected post-state, the prior–updated relation captures the revision induced by new context, and the updated–target relation reveals what remains unexplained after that revision.

LEVIR-CD+ · scene 881

(I<sub>2</sub>; V<sub>2</sub>) — (I₂; V₂)

LEVIR-CD+ · scene 947

SYSU-CD · 01030

SYSU-CD · 00316

SYSU-CD · 01081

SYSU-CD · 01818

CDD · 00013

CDD · 00722

Figure 10: The same visualisation on further test pairs from LEVIR-CD+ (full scenes), SYSU-CD, and CDD. Latent states are shown as the first three principal components of a joint PCA of the latent states of each example; evidence maps are cosine distances computed with a single fixed-seed mask, clipped to the 2nd–98th percentile. Green contours mark the ground-truth change. LEVIR-CD+ scenes are processed as sixteen independent 256×256 crops, so crop boundaries remain visible in the latent and evidence maps.

Implementation Details

Initialisation	I-JEPA ViT-H/16 (ImageNet-1K, 448×448, 300 epochs) for the context encoder, target encoder, and predictor
Target encoder	Exponential moving average of the context encoder, momentum increased linearly from 0.996 to 1.0
Optimisation	AdamW, weight decay 0.05, gradient clipping 1.0, batch size 12; peak learning rates 3×10⁻⁴ (new modules), 1×10⁻⁴ (predictor), 1×10⁻⁵ (context encoder); linear warm-up over 15,000 samples followed by cosine decay
Schedule	30,000 iterations with L_JEPA only, followed by 125,000 iterations of joint optimisation
Losses	L_cd = L_BCE + L_Tv + λ_LovL_Lov; λ_prior = 0.25, λ_JEPA = 1.0, λ_Lov = 0.5, Tversky α = β = 0.5
Inference	Hidden post-change ratio 0.7; logits averaged over five fixed-seed block masks and the original, horizontally flipped, and vertically flipped inputs
Hardware	Single NVIDIA RTX A6000 (48 GB); 256×256 bitemporal RGB inputs

Table 3: Training and inference settings.

Dataset	Released sample size	# Train	# Test
LEVIR-CD+	1024×1024	10,192	5,568
SYSU-CD	256×256	16,000	4,000
CDD	256×256	12,998	3,000

Table 4: Dataset partitions. All samples presented to the network are 256×256 bitemporal pairs; LEVIR-CD+ images are partitioned into a 4×4 grid of non-overlapping crops.

Conclusion

We introduced ChangeJEPA, a bitemporal JEPA framework that complements direct comparison with latent post-state prediction. ChangeJEPA forms a prior hypothesis from the pre-change image and updates it using partial post-change context. Comparisons among the pre-change, observed, and predicted states provide complementary evidence of observed difference, prior expectation, contextual update, and residual inconsistency.

Across all benchmarks, ChangeJEPA consistently outperforms its decoder-matched Siamese baseline and achieves the highest IoU, F1-score, and OA on SYSU-CD, LEVIR-CD+, and CDD. The ablations confirm that direct bitemporal evidence remains essential, while the prior and updated predictions contribute distinct and complementary information. These results suggest that latent post-state prediction provides a principled complement to conventional Siamese comparison.

ChangeJEPA advances change detection by predicting before comparing: it establishes a common post-change reference, interprets how observations revise expectations, and uses both explained change and unexpected discrepancies as evidence.

BibTeX

@misc{wijenayake2026changejepapredictcompareremote,
      title={ChangeJEPA: Predict Before You Compare for Remote Sensing Change Detection},
      author={Buddhi Wijenayake and Dhananjaya Jayasundara and Vishal M. Patel},
      year={2026},
      eprint={2611.05570},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2611.05570},
}

ChangeJEPA

Predict Before You Compare forRemote Sensing Change Detection

Predict the scene. Interpret the differences. Reveal the change.