Remote sensing change detection usually relies on large collections of temporally aligned, pixel-level annotations, which limits its scalability across regions, imaging platforms and application domains. Recent zero-shot approaches reduce the need for supervision, but they often localize fine-grained changes poorly because their features are coarse or inconsistent across time. We present SpotDiff, a training-free, plug-and-play framework for zero-shot change detection built by composing complementary vision foundation models. SpotDiff uses DINOv3 for semantically robust features, AnyUp to reconstruct them at pixel resolution, and SAM 3 for text-guided region-of-interest proposals, which an inference-time multi-scale strategy (SWIMS) makes reliable on small objects. Change is then measured by comparing the reconstructed features of the two dates inside the proposed regions, with no learnable components, fine-tuning or change labels. To our knowledge, SpotDiff is the first change detection method to use a universal feature upsampler, which alone adds 13.8 F1 over bilinear upsampling of the same features. On LEVIR-CD, SpotDiff reaches 82.46 F1, 2.06 points above the strongest prior training-free method, and on the multi-class SECOND benchmark it improves building change detection by 10.3 IoU and the six-class average by 2.2 IoU. Qualitative results on PlanetScope (3 m) and Sentinel-2 (10 m) imagery show that the same pipeline transfers across resolutions without modification. Our results indicate that composing modern foundation models is an effective route to scalable, label-free change detection.
Figures & tables
Figure 1: Overview of SpotDiff. The Similarity Module encodes both images with a frozen vision foundation model, reconstructs pixel-resolution features with a universal upsampler and computes their per-pixel cosine similarity. The Region of Interest Proposal Module segments instances of the prompted category in each image with multi-scale inference (SWIMS) and fuses them across time. The Change Detection Module marks a candidate instance as changed when its mean similarity falls below a threshold. No component is trained.
Method
Params (M)
TFLOPs
Time (s)
F1
OmniOVCD [ 35 ]
840
–
–
80.40
AdaptOVCD [ 14 ]
1,365
–
–
68.00
SpotDiff (ViT-L)
1,144
134.5
9.6
80.10
SpotDiff (ViT-7B)
7,557
311.4
≈ 25
82.46
Table 1: Compute for one 1024×1024 pair; F1 on LEVIR-CD (%). Params are frozen foundation-model weights (none are trained). FLOPs counted with PyTorch’s FlopCounterMode (1 multiply-add = 2 FLOPs). SpotDiff: SAM 3 840M, AnyUp 0.9M, DINOv3 303M (ViT-L) or 6,716M (ViT-7B); TFLOPs per pair = encoder 8.3 / 132.2, AnyUp 24.3 / 77.3, SWIMS 101.9 (91.7 of it at 2× ). OmniOVCD: SAM 3 weights; AdaptOVCD: sum of the parameter counts in its paper. Time on one H100. “–”: not reported.
Method
LEVIR-CD
WHU-CD
S2Looking
IoU
F1
IoU
F1
IoU
F1
PCA-KM [ 5 ]
4.8
9.1
5.4
10.2
–
–
CNN-CD [ 16 ]
7.0
13.1
4.9
9.4
–
–
DSFA [ 15 ]
4.3
8.2
4.1
7.8
–
–
DCVA [ 27 ]
7.6
14.1
10.9
19.6
–
–
GMCD [ 31 ]
6.1
11.6
10.9
19.7
–
–
Table 2: Training-free building change detection (IoU / F1, %). Baselines from [ 35 ] except AdaptOVCD and OpenDPR (from their papers); ∗ re-implemented there under identical settings; “–”: not reported. SpotDiff uses SWIMS {1×,2×} , τ=0.6 and the prompt “roof”. † Evaluated on 1024×1024 test tiles rather than the 256×256 patches of [ 35 ] .
Method
Building
Tree
Water
Low veg.
Surface
Playground
Class avg.
IoU
F1
IoU
F1
IoU
F1
IoU
F1
IoU
F1
IoU
F1
IoU
F1
SAM-DINOv2-SegEarth-OV ∗ [ 23 ]
38.8
56.0
15.6
27.0
15.3
26.5
21.3
35.1
26.7
42.1
22.6
36.8
23.4
37.3
SAM2-DINOv2-SegEarth-OV ∗ [ 23 ]
36.3
53.3
15.8
27.3
14.6
25.5
20.1
33.4
19.6
32.7
24.6
39.5
21.8
35.3
APE-/-DINO ∗ [ 23 ]
29.1
45.1
9.7
17.7
12.3
22.0
–
–
–
–
25.6
40.8
12.8
21.0
APE-/-DINOv2 ∗ [ 23 ]
31.9
48.3
10.6
19.2
12.2
21.7
–
–
–
–
25.0
40.0
13.3
21.5
OmniOVCD [ 35 ]
45.2
62.3
16.7
28.6
21.2
35.0
24.5
39.3
27.7
43.4
27.0
42.4
27.1
41.8
Table 3: Per-class change detection on SECOND (IoU / F1, %), each class evaluated as binary change following [ 23 ] . Baselines from [ 35 ] except AdaptOVCD and OpenDPR (from their papers; OpenDPR in its multi-class inference setting); ∗ re-implemented there; “–” denotes a score close to 0; AdaptOVCD’s class average is the mean of its six per-class values, computed by us. SpotDiff’s class average is computed from one complete run of all six classes and differs slightly from the mean of the row (29.80 IoU) because of run-to-run variation in SAM 3 proposals.
Figure 3: Qualitative multi-class results on SECOND (selected test pairs, image IDs on the left). Columns: image at t0 , image at t1 , ground truth, SpotDiff prediction, OmniOVCD prediction. Each class is detected as binary change with its own prompts; colours indicate the changed class (ground truth: class at t1 ; black: no change). Where several classes fire on a pixel, SpotDiff shows the class of the instance with the lowest similarity; OmniOVCD exposes no instance score, so its overlaps follow a fixed priority (playground, water, building, tree, low vegetation, surface).
Figure 4: Cross-resolution results of the class-agnostic variant ( τ=0.7 ) on medium-resolution imagery, using the same models and settings as on sub-meter data. Magenta: pixels with cosine similarity below τ (changed). (a) PlanetScope at 3 m: deforestation, urban and port changes. (b) Sentinel-2 RGB at 10 m: building, road and land-clearing changes.
Upsampler
F1
IoU
Prec.
Rec.
Nearest neighbor
66.88
54.72
57.79
90.44
Bilinear
68.69
56.19
55.20
90.94
AnyUp
82.46
70.15
74.70
92.00
Table 4: Upsampler ablation on LEVIR-CD (%).
Backbone
Scales
0.5
0.6
0.65
0.7
0.8
ViT-7B
1×
55.97
61.09
61.16
59.43
55.05
1× + 2×
78.54
82.46
82.11
80.12
76.42
1× + 2× + 4×
76.29
77.36
76.63
74.12
69.53
ViT-L
1×
52.19
58.16
59.75
59.95
56.87
1× + 2×
73.64
80.10
81.22
81.29
77.91
1× + 2× + 4×
70.51
75.29
75.90
75.45
71.33
Table 5: SWIMS scales and threshold τ on LEVIR-CD (F1, %). Columns are values of τ .
Figure 5: SWIMS vs. single-scale inference on LEVIR-CD. SAM 3 instances at 1× (top) and with SWIMS {1×,2×} (bottom).
Figure 6: Sensitivity on LEVIR-CD. (a) F1 against the similarity threshold τ for each SWIMS scale setting and encoder ( Tab. 5 ). (b) F1 for four synonymous SAM 3 prompts ( Tab. 6 ).
Remote sensing (RS) change detection is essential for interpreting surface dynamics. Semantic change detection (SCD) further enables pixel-level understanding of multi-class transitions, yet remains sensitive to pseudo-changes induced by imaging conditions. Recent RS foundation models extract semantically consistent features across temporal and environmental variations, which is critical for mitigating pseudo-changes. However, existing SCD methods are often rigid and backbone-specific, lacking the flexibility to integrate diverse multi-scale features from emerging foundation models. To this end, we introduce a modular Cascaded Gated Decoder (CG-Decoder) that bridges various backbones and SCD tasks, processing multi-scale features in a coarse-to-fine manner while enabling adaptive change extraction. Building upon the RS foundation model PerA, we present PerASCD, a unified SCD framework. We further propose a Soft Semantic Consistency Loss (SSCLoss) to mitigate numerical instability in mixed-precision training. Extensive experiments on SECOND and LandsatSCD show that PerASCD achieves new state-of-the-art Sek scores (26.11% and 65.21%), surpassing the previous best by 0.61% and 4.95%, respectively. It also demonstrates exceptional data efficiency (outperforming the full-data baseline with 50% data), seamless cross-backbone generalization, and enhanced interpretability. Our approach maintains robust semantic consistency under radiometric variations, providing a reliable SCD solution. Code: https://github.com/SathShen/PerASCD.git.
Hengtong Shen, Li Yan, Hong Xie +5
The authors are with the School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, China
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
Buddhi Wijenayake, Dhananjaya Jayasundara, Vishal M. Patel
Unsupervised remote sensing change detection (UCD) aims to localise changes between two images of the same region without relying on labelled training data. Most recent approaches either use a frozen foundation model in a training-free manner or train with synthetic changes generated in pixel space. Both strategies inherently rely on predefined assumptions about change types, typically introduced through handcrafted rules, external datasets, or auxiliary generative models. Due to these assumptions, such methods fail to generalise beyond a few change types, limiting their real-world usage, especially in rare or complex scenarios. To address this, we propose MaSoN (Make Some Noise), an end-to-end UCD framework that synthesises diverse changes directly in the latent feature space during training. It generates changes dynamically estimated from feature statistics of the target data, enabling diverse yet data-driven variation aligned with the target domain. Since synthesis happens in latent space, it also easily extends to new modalities, such as SAR and multispectral data. MaSoN generalises strongly across diverse change types and improves the average F1 score across five benchmarks by 14.1 percentage points. Project page: https://blaz-r.github.io/mason_ucd/
Blaž Rolih, Matic Fučka, Filip Wolf +1
University of Ljubljana, Faculty of Computer and Information Science, Slovenia