Change Detection & Captioning (CDC) aims to jointly localize changed regions and describe their semantic content from bi-temporal remote sensing images. While recent approaches have demonstrated the effectiveness of unified modeling, they primarily rely on shared representations or jointly optimized architectures, leaving task-specific communication under heterogeneous supervision less explored. However, detection and captioning differ fundamentally in their supervision granularity, leading to heterogeneous representations with uneven reliability across tasks and spatial locations. In this paper, we reformulate CDC as a selective information routing problem between heterogeneous task branches, where the key challenge is to determine when and where information should be transferred. To this end, we propose CoDeRNet, a unified framework that preserves task-specific representations while enabling bidirectional exchange of complementary information. CoDeRNet introduces a Confidence-Decomposed Routing (CoDeR) mechanism, which decomposes the routing condition into receiver-side need for support and sender-side reliability cues to guide selective transfer. Experiments on LEVIR-MCI and WHU-CDC show that CoDeRNet achieves strong overall CDC performance, particularly improving caption generation while preserving competitive change localization. Extensive analyses further show that the benefits arise not merely from joint learning or generic cross-task communication, but from explicitly modeling selective cross-task information routing. These results highlight the importance of conditional information transfer for unified CDC under heterogeneous supervision.
Figures & tables
Figure 1: Overview of CoDeRNet. Shared change features are transformed into task-specific detection and captioning features, followed by Confidence-Decomposed Routing (CoDeR) for selective bidirectional transfer before task-specific decoding.
Task
Method
Metrics
mIoU
BLEU-1
BLEU-2
BLEU-3
BLEU-4
METEOR
ROUGE-L
CIDEr-D
Change Detection
SNUNet [ 12 ]
82.53
-
-
-
-
-
-
-
ChangeFormer [ 1 ]
84.35
-
-
-
-
-
-
-
TFI-GR [ 19 ]
79.34
-
-
-
-
-
-
-
A2Net [ 18 ]
80.82
-
-
-
-
-
-
-
HANet [ 13 ]
71.23
-
-
-
-
-
-
-
Table 1: Comparison on LEVIR-MCI [ 22 ] . “Change Detection” and “Change Captioning” denote single-task models trained for only one task, whereas “Change Detection & Captioning” denotes unified models that perform both tasks jointly within a single framework. Bold and underlined values indicate the best and second-best results, respectively.
Task
Method
Metrics
IoU
BLEU-1
BLEU-2
BLEU-3
BLEU-4
METEOR
ROUGE-L
CIDEr-D
Change Detection
SNUNet [ 12 ]
74.81
-
-
-
-
-
-
-
ChangeFormer [ 1 ]
77.88
-
-
-
-
-
-
-
TFI-GR [ 19 ]
77.60
-
-
-
-
-
-
-
A2Net [ 18 ]
80.43
-
-
-
-
-
-
-
HANet [ 13 ]
79.95
-
-
-
-
-
-
-
Table 2: Comparison on WHU-CDC [ 35 ] . “Change Detection” and “Change Captioning” denote single-task models trained for only one task, whereas “Change Detection & Captioning” denotes unified models that perform both tasks jointly within a single framework. Bold and underlined values indicate the best and second-best results, respectively.
Figure 2: Qualitative comparison on LEVIR-MCI. In the captions, building-related words (e.g., house , building ) and road-related words (e.g., road ) are highlighted using colors consistent with the corresponding mask classes.
Seg
Cap
CoDeR
mIoU
BLEU-4
METEOR
ROUGE-L
CIDEr-D
✓
-
-
87.03
-
-
-
-
-
✓
-
-
63.18
40.21
75.27
137.41
✓
✓
-
87.67
64.41
39.93
74.98
136.18
✓
✓
✓
87.79
66.27
41.08
76.36
141.79
Table 3: Ablation studies on LEVIR-MCI [ 22 ] . (a) Effect of joint training and explicit interaction, comparing single-task training, joint training without CoDeR, and joint training with CoDeR. (b) Effect of CoDeR directionality, comparing two single-direction settings with the bidirectional design. Bold indicates the best result in each column.
Mask
mIoU
BLEU-4
METEOR
ROUGE-L
CIDEr-D
Sender reliability (Cs)
87.60
65.12
40.49
75.73
139.57
Receiver need (1−Ct)
87.65
64.65
40.81
75.94
141.02
Single gate
87.59
65.46
40.27
75.54
138.19
Decomposed routing ((1−Ct)⋅Cs)
87.79
66.27
41.08
76.36
141.79
Table 4: Ablation studies on LEVIR-MCI [ 22 ] for analyzing the effectiveness of CoDeR. (a) Routing mask decomposition ablation compares CoDeR with reduced variants and a single-gate design. (b) Interaction operator ablation compares CoDeR with representative cross-task communication operators under the same framework. Bold and underline indicate the best and second-best results in each column, respectively.
Model
Factual grounding
Cross-output consistency
MCINet [ 22 ]
3.520±1.38
3.632±1.25
Ours (w/o CoDeR)
3.645±1.22
3.782±1.09
Ours (w/ CoDeR)
3.658±1.27
3.813±1.13
Table 5: Blind human evaluation on the change-containing subset of LEVIR-MCI. Values are mean ± standard deviation on a five-point scale, with 840 ratings per system for each criterion across 660 unique image pairs. Higher scores indicate better factual grounding or cross-output consistency. Bold indicates the highest mean in each column.
Table 6: Ablation of the CoDeR routing scale on LEVIR-MCI [ 22 ] . Each subtable reports performance when same-scale routing is applied at different receiving scales in one direction, while the opposite direction is fixed to its default setting. Bold indicates the best result in each column.
Backbone
CoDeR
mIoU
BLEU-4
MET.
R-L
CID.
ConvNeXt-B [ 27 ]
×
87.67
64.41
39.93
74.98
136.18
✓
87.79
66.27
41.08
76.36
141.79
SwinV2-B [ 26 ]
×
87.49
63.55
40.24
75.26
136.93
✓
87.34
64.45
40.30
75.31
138.20
MiT-B5 [ 42 ]
×
87.74
63.92
40.15
75.49
137.54
✓
87.78
65.38
40.54
75.54
140.28
Table 7: Robustness of CoDeRNet across backbone choices on LEVIR-MCI [ 22 ] . (a) Backbone-family ablation: performance changes with and without CoDeR across different backbones. (b) Backbone-size ablation: performance changes with and without CoDeR across ConvNeXt variants from Tiny to Large. Bold indicates the best result in each column. In the CoDeRNet rows, colored values indicate improvements ( gain ) or degradations ( drop ) relative to the corresponding baseline. B4, MET., R-L, and CID. denote BLEU-4, METEOR, ROUGE-L, and CIDEr-D, respectively.
Type
Setting
mIoU
BLEU-4
METEOR
ROUGE-L
CIDEr-D
Multi
both (all↔all)
87.66
65.81
40.18
75.93
138.83
seg only (all↔1/16)
87.75
65.44
41.03
76.22
142.18
cap only (1/8↔all)
87.86
64.56
40.83
75.93
140.32
Single (Ours)
(1/8↔1/16)
87.79
66.27
41.08
76.36
141.79
Table 8: Ablation studies on LEVIR-MCI [ 22 ] . (a) Single-scale versus multi-scale CoDeR application. The default design applies CoDeR at one receiving scale per direction (seg ← cap at 1/8 , cap ← seg at 1/16 ), while the multi-scale variants apply independent CoDeR blocks at all four scales for the indicated direction(s). (b) Region-token pooling size ablation for caption decoding. Bold indicates the best result in each column.
Aug.
mIoU
BLEU-4
METEOR
ROUGE-L
CIDEr-D
-
86.61
64.51
39.44
74.72
136.06
G
86.67
64.45
40.64
75.97
139.21
G+R
87.79
66.27
41.08
76.36
141.79
Table 9: Additional ablation studies on LEVIR-MCI [ 22 ] . (a) Effect of augmentation type, where G denotes geometric augmentation and R denotes directional caption remapping. (b) Effect of change fusion composition, where F is formed from temporal features at t0 and t1 together with signed and magnitude change cues Δ and ∣Δ∣ . Bold indicates the best result in each column.
Transform
Category
Rule / Example
Remapping rules
H-Flip
Single-level
left↔right
Phrase-level
upper left↔upper right
Phrase-level
lower left↔lower right
Phrase-level
top left↔top right
Phrase-level
bottom left↔bottom right
Table 10: Directional caption remapping under geometric augmentation on LEVIR-MCI. To preserve caption–image consistency, direction-sensitive expressions are remapped according to the applied geometric transform. Two-word directional phrases are matched prior to single-word directional levels.
Seg
Cap
CoDeR
mIoU
cIoU
✓
×
×
87.03
82.05
✓
✓
×
87.67
82.97
✓
✓
✓
87.79
83.07
Table 11: Additional detection evaluation. (a) Multi-class mIoU and change-class IoU (cIoU) on LEVIR-MCI [ 22 ] . Seg and Cap indicate change detection and change captioning training, respectively. cIoU is computed by merging building and road into a single change class. Bold indicates the best result in each metric column. (b) Image-level paired bootstrap analysis. Δ denotes the full-test-set performance difference between CoDeRNet and MCINet. Differences and 95% confidence intervals are expressed in percentage points. Positive values favor CoDeRNet.
Figure 3: Qualitative comparison on LEVIR-MCI. In the captions, building-related words (e.g., house , building ) and road-related words (e.g., road ) are highlighted using colors consistent with the corresponding mask classes.
Figure 4: Qualitative comparison on WHU-CDC. In the captions, building-related words (e.g., house , building ) are highlighted using colors consistent with the corresponding mask classes.
Understanding and localizing subtle changes between paired images is critical for tasks such as surveillance and image editing. However, traditional Image Change Captioning (ICC) methods lack spatial grounding, limiting their precision. We introduce Image Change Captioning and Segmentation (ICCS), a new multimodal task that jointly requires structured change description and pixel-level localization. To address ICCS, we propose the Change-aware Captioning and Reasoning Chain (CCRC), a dual-chain framework that decouples semantic reasoning from spatial segmentation. The first chain, Chain-of-Change-Captioning (CCC), enhances fine-grained change perception via a visual fusion module based on Multi-Head Change-aware Attention inserted between the visual and language components of a Multimodal Large Language Model (MLLM). CCC also determines whether a change is segmentable. If not, it alone generates the caption. Otherwise, the second chain, Chain-of-Change-Segmenting (CCS), is activated, leveraging spatial priors from CCC and refining masks with a Change-aware Token Refiner for accurate boundary localization. We evaluate CCRC on both synthetic and real-world change detection benchmarks with pixel-level supervision. Experiments show CCRC achieves state-of-the-art performance.
Jinhong Hu, Xiaoping Wang, Shuyin Huang +3
College of Computer Science and Electronic Engineering, Hunan University, China · Guangxi Minzu University, China · National University of Defense Technology, China
Remote sensing change detection (RSCD) is essential for monitoring land-cover changes and urban development. However, most methods demand pixel-level change masks, which are costly and time-consuming to annotate. Weakly supervised methods reduce this cost by using image-level change labels. Yet these labels indicate only whether a change occurs, leaving models to recover the location of the change and semantic meaning through additional and complex mechanisms. This missing information can be supplied directly by change captions, which describe what changes, what it becomes, and where it occurs. Therefore, we introduce change-caption-guided RSCD, using change captions as the sole task-specific supervision to learn change masks without manually annotated change masks. Our framework has two components: a caption-driven generation pipeline that produces bi-temporal remote sensing image pairs at scale with controlled changes matching each caption, and a change detector guided by the caption's transition semantics. The detector uses our Semantic-Appearance Agreement Framework (SAAF) to combine caption-grounded semantic responses with RGB differences for change localization, while text conditioning guides dense prediction. Experiments on our newly constructed Flair-RSGen dataset and WHU-CDC show that SAAF outperforms the closest reproduced limited-supervision baselines in macro-averaged IoU and F1 under the evaluated protocols. Code is publicly available at https://github.com/qianyuancs/SAAF.
Yuan Qian, Jie Ma
School of Information Science and Technology, Beijing Foreign Studies University, Beijing 100089, China
Remote sensing change detection (CD) from bi-temporal imagery is critical for applications such as urban monitoring, disaster assessment, and environmental management, yet robust localization remains challenging under sparse changes, noisy labels, and appearance variations. In this paper, we propose Context Sampling Attention (CoSA), a lightweight decoder-side refinement module that explicitly leverages bi-temporal feature correlation as a control signal for adaptive change-aware feature enhancement. This differs from conventional attention mechanisms that rely on implicit feature weighting without explicit temporal control. In the implemented FC-Siam setting, CoSA computes normalized same-location cross-correlation between paired decoder features, converts low correlation into a change gate, and injects the resulting gated residual at native 1/8 and 1/16 feature scales through learnable residual scaling. This design enables effective discrimination between stable and ambiguous regions without relying on computationally expensive global attention. Extensive experiments on four benchmark datasets (LEVIR-CD, S2Looking, DSIFN, and CLCD) demonstrate consistent improvements over strong baselines, achieving 1.5-2.6% gains in changed-class F1 while introducing negligible parameter overhead. Ablation studies confirm that multiscale placement and learnable residual gating are both important for peak performance. These results indicate that CoSA establishes a practical and effective refinement paradigm for enhancing temporal discriminability in Siamese change detection frameworks.
Abdirashid Omar, Jonghyuk Park
Department of Data Science, Graduate School of Kookmin University Seoul 02707, Republic of Korea