DS-BiCA-Net: Pair-Preserving Cross-Temporal Modeling for Remote Sensing Change Captioning

Licheng Zhang; Naveed Akhtar; Siew-Kei Lam

cs.CVOct 23, 2026

DS-BiCA-Net: Pair-Preserving Cross-Temporal Modeling for Remote Sensing Change Captioning

Authors: Licheng Zhang, Naveed Akhtar, Siew-Kei Lam

Organizations: the School of Computing and Information Systems, the University of Melbourne, Victoria 3010, Australia · Nanyang Technological University, Singapore

Abstract

Remote sensing image change captioning (RSICC) describes in natural language the semantic changes between two images of the same scene captured at different times. Previous methods commonly merge the pair during repeated processing or combine intra-image and cross-image relations in one operation, weakening source identity and leaving the two temporal directions implicit. Without separate modules for intra-image and cross-temporal processing and a separate path for each direction, the contribution of each part cannot be measured on its own, so it remains unclear which part matters most, and the design of later models has no such evidence to build on. To address these issues, we propose DS-BiCA-Net, which assigns the two kinds of processing to separate modules and keeps the two images as separate streams throughout the encoder. A windowed state-space module models context within each image, and a bidirectional cross-attention module compares the two images along two directed paths that share their projections. The two modules use the same pre-normalisation residual architecture and differ only in their mixers, and the two streams are fused only in the aggregation head, so each module and each direction can be removed and its contribution measured on its own. Controlled experiments on LEVIR-CC and Dubai-CC examine operator interaction, temporal direction and module order, showing that intra-image refinement carries no benefit alone and becomes beneficial only when followed by bidirectional cross-attention, an order that outperforms the reverse at matched parameter count and FLOPs. The bidirectional module also outperforms either temporal direction alone. Our method reaches an S_m^* of 82.65 on LEVIR-CC, the highest among the published results compared, and 66.72 on Dubai-CC without geometric augmentation, the second-highest among the published results compared.

Figures & tables

Fig. 1: Comparison interfaces of Chg2Cap [ 10 ] , RSCaMa [ 18 ] , TISDNet [ 12 ] and DS-BiCA-Net.

X_{0}

and

X_{1}

are the features of the earlier and the later image of the pair, across which the change is described. Each block is repeated

N

times.

Method	Self-image mixing	Pair output	Directed paths	Multi-layer aggregation
Chg2Cap [ 10 ]	✓	✓	✗	✗
RSCaMa [ 18 ]	✓	✓	✗	✗
TISDNet [ 12 ]	✗	✓	✓	✓
DAE [ 14 ]	✓	✓	✗	✓
RMNet [ 20 ]	✓	✓	✗	✗
MTH-Net [ 32 ]	✓	✓	✗	✗

TABLE I: How existing change-aware encoders process and compare the two images.

Fig. 2: Overview of DS-BiCA-Net. A shared ResNet101 extracts features of the earlier image

I_{0}

and the later image

I_{1}

at three scales. At each scale, a cascade of DS-BiCA blocks first refines each image with its own spatial context (DBSS) and then lets each image attend to the other in both directions (BiCA), keeping the two images as separate streams. The aggregation head fuses the outputs of all blocks into one representation per scale, and the decoder reads the three scales through gated cross-attention to generate the caption word by word.

Panel of Figure 4 — Fig. 3: Pair-preserving DS-BiCA encoder. Here $\oplus$ denotes residual addition and $\odot$ an elementwise product. In panels (c) and (d), $H$ , $W$ , and $C$ are the feature height, width, and channel count, $L=HW$ is the token count, and $x\in\mathbb{R}^{L\times C}$ is the input sequence. The panels share one module structure and differ only in the mixer, and each footer names the equation that defines a box. Panel (c) is the DBSS module shared by $t=0$ and $t=1$ , a pre-norm residual pair around a selective-scan mixer and an MLP. Panel (d) is the BiCA module, a pre-norm cross-attention whose query $F_{t}$ comes from one stream and keys and values $F_{1-t}$ come from the other, followed by an MLP residual. The layer scales $\gamma_{1}$ and $\gamma_{2}$ are those of Eq. ( 7 ).

Fig. 4:

S_{m}^{*}

of the sixteen methods that appear on both benchmarks, with our model at protocol 1 on Dubai-CC. The dashed line is the least-squares fit.

Attention-based comparison
Method	B-1	B-2	B-3	B-4	M	R-L	C	$S_{m}^{*}$
CD4C [ 26 ]	$86.74$	$78.71$	$71.48$	$65.41$	$40.56$	$75.72$	$138.00$	$79.92$
DetACC [ 27 ]	$87.26$	$79.16$	$71.74$	$65.56$	$40.65$	$75.97$	$139.99$	$80.54$
DSPG [ 37 ]	$86.49$	$78.81$	$72.21$	$66.71$	$41.61$	$76.81$	$143.34$	$82.12$
SGD-RSCCN [ 40 ]	$84.17$	$75.16$	$68.05$	$62.48$	$39.18$	$74.24$	$136.27$	$78.04$
RSICCformer [ 6 ]	$84.72$	$76.27$	$68.87$	$62.77$	$39.61$	$74.12$	$134.12$	$77.66$

TABLE II: Performance comparison on the LEVIR-CC dataset.

Attention-based comparison
Method	B-1	B-2	B-3	B-4	M	R-L	C	$S_{m}^{*}$
TACC [ 30 ]	$72.17$	$58.65$	$48.24$	$38.59$	$28.17$	$58.93$	$95.45$	$55.29$
CD4C [ 26 ]	$74.13$	$63.11$	$53.13$	$43.55$	$30.27$	$60.25$	$108.66$	$60.68$
DSPG [ 37 ]	$72.34$	$60.52$	$50.30$	$40.65$	$29.40$	$62.66$	$100.42$	$58.28$
SGD-RSCCN [ 40 ]	77.15	66.80	58.07	50.27	34.07	$65.88$	$118.42$	67.16
RSICCformer [ 6 ]	$67.92$	$53.61$	$41.37$	$31.28$	$25.41$	$51.96$	$66.54$	$43.80$

TABLE III: Performance comparison on the Dubai-CC dataset. Bold and underline rank the published methods together with protocol 1. Protocols 2 and 3 use geometric augmentation and are listed below the last rule for reference only.

Fig. 5: Accuracy and capacity trade-off on LEVIR-CC.

DS-BiCA-Net ( Ours )
Method	Pair	Steps	Para. (M)	Mean $\pm$ std. (ms)	FPS
SEN [ 16 ]	$256^{2}$	$52$	$26.48$	$26.57\pm 1$ .24)	$37.6$
Chg2Cap [ 10 ]	$256^{2}$	$41$	$328.89$	$38.63\pm 1$ .26)	$25.9$
ICT-Net [ 11 ]	$256^{2}$	$52$	$96.64$	$53.04\pm 0$ .70)	$18.9$
RSCaMa [ 18 ]	$224^{2}$	$42$	$898.93$	$31.05\pm 1$ .06)	$32.2$
RMNet [ 20 ]	$224^{2}$	$52$	$129.18$	$59.90\pm 17$ .01)	$16.7$

TABLE IV: Parameter-count and end-to-end latency comparison. “Pair” is the input resolution and “FPS” is

1000/\text{mean}

Panel of Figure 10 — Fig. 6: Qualitative comparison on the LEVIR-CC test set.

Panel of Figure 11 — Fig. 7: Dubai-CC qualitative comparison across the three training protocols.

Panel of Figure 12 — Fig. 8: Word-level decoder cross-attention decomposed by scale.

Error type	LEVIR-CC	Dubai-CC
Small-change miss	89 (4.61%)	5 (3.33%)
Wrong object	7 (0.36%)	1 (0.67%)
Wrong object position	0 (0.00%)	0 (0.00%)
Wrong count	40 (2.07%)	0 (0.00%)
Pseudo-change	28 (1.45%)	3 (2.00%)
Unsupported-object hallucination	79 (4.10%)	4 (2.67%)

TABLE V: Caption-level failure analysis on LEVIR-CC and Dubai-CC.

(a) LEVIR-CC
Setting and model	B-1	B-2	B-3	B-4	M	R-L	C	$S_{m}^{*}$	Para. (M)	FLOPs (G)
Identity	$87.05$	$79.02$	$72.03$	$66.30$	$41.59$	$76.79$	$143.46$	$82.04$	$53.78$	$26.70$
DBSS-only	$86.92$	$79.20$	$72.56$	$67.23$	$41.24$	$76.18$	$141.59$	$81.56$	$56.43$	$33.82$
BiCA-only	87.53	79.45	$72.13$	$65.94$	$41.27$	77.20	144.45	$82.21$	$56.93$	$37.45$
UniCA $1\!\leftarrow\!0$	$85.32$	$77.63$	$71.16$	$65.94$	$40.50$	$75.45$	$139.91$	$80.45$	$59.58$	$39.20$
UniCA $0\!\leftarrow\!1$	$86.48$	$79.04$	$72.57$	$67.23$	$41.19$	$76.29$	$142.76$	$81.87$	$59.58$	$39.20$

TABLE VI: Ablation studies. Metric abbreviations follow Table II .

Fig. 9: Factorial and direction controls from Table VI (a) and (b). Bars show complete-test scores with marginal 95% bootstrap intervals. Paired tests appear in Table VII .

Fig. 10: Caption quality under the two encoder scaling axes.

Contrast	$\Delta S_{m}^{*}$	95% CI	$p$	Holm $p$
LEVIR-CC
Identity	$0.61$	$[-0.48,1.57]$	$0.230$	$0.919$
DBSS-only	$1.09$	$[0.06,2.15]$	$0.038$	$0.252$
BiCA-only	$0.44$	$[-0.54,1.55]$	$0.401$	$0.919$
UniCA $1\!\leftarrow\!0$	2.20	[1.05, 3.42]	0.001	0.010
UniCA $0\!\leftarrow\!1$	$0.78$	$[-0.24,1.81]$	$0.117$	$0.584$

TABLE VII: Paired bootstrap comparisons against the single-block configuration, the best model on LEVIR-CC and on Dubai-CC protocol 1.

\Delta S_{m}^{*}

is the difference of the complete-test-set scores in Table VI , while the interval and

p

come from the bootstrap replicates. The last block reports the

2\times 2

factorial interaction, which is exploratory and outside the corrected family.

Model	BiCA $\gamma$	DBSS $\gamma$	$\times$ init	$\Delta$ input (%)
LEVIR-CC, $S_{m}^{*}$ $82.65$	$9.96\text{\times}{10}^{-3}$	$3.32\text{\times}{10}^{-3}$	$664$	$20.4$
Dubai-CC, $S_{m}^{*}$ $72.13$	$3.43\text{\times}{10}^{-3}$	$1.57\text{\times}{10}^{-3}$	$219$	$4.0$

TABLE VIII: Encoder activity in the two models.

Remote Sensing Image Change Captioning

Remote sensing image change captioning describes meaningful changes between satellite or aerial images in natural language.

See all 13 papers

May 14, 2026

HiSem: Hierarchical Semantic Disentangling for Remote Sensing Image Change Captioning

Remote sensing image change captioning (RSICC) aims to achieve high-level semantic understanding of genuine changes occurring between bi-temporal images. Despite notable progress, existing methods are fundamentally limited by a shared modeling assumption: changed and unchanged image pairs, which have intrinsically different semantic granularities, are processed under a unified modeling strategy. This modeling inconsistency leads to semantic entanglement between coarse-grained change existence judgment and fine-grained semantic understanding.To address the above limitation, we propose a novel hierarchical semantic disentangling network (HiSem) that explicitly disentangles semantic representations of different granularities. Specifically, we first introduce the Bidirectional Differential Attention Modulation (BDAM) module that leverages discrepancy-aware attention to enhance cross-temporal interactions, thereby amplifying true change signals while suppressing irrelevant variations. Building upon this, we design a Hierarchical Adaptive Semantic Disentanglement (HASD) module that performs adaptive routing at two hierarchical levels: a coarse-grained image-level routing mechanism distinguishes changed and unchanged image pairs, while a fine-grained token-level Mixture-of-Experts (MoE) block models diverse and heterogeneous change semantics for changed samples. Extensive experiments on two benchmark datasets demonstrate that HiSem outperfoms previous methods, achieving a significant improvement of +7.52% BLEU-4 on the WHU-CDC dataset. More importantly, our approach provides a structured perspective for RSICC by explicitly aligning model design with the intrinsic semantic heterogeneity of bi-temporal scenes. The code will be available at https://github.com/Man-Wang-star/HiSem

Man Wang, Chenyang Liu, Wenjun Li +5

College of Computer Science, Inner Mongolia University, Hohhot 010021, China · Research Center for Spatiotemporal Intelligence, Inner Mongolia University, Hohhot 010021, China · Department of Aerospace Intelligent Science and Technology, School of Astronautics, Beihang University, Beijing 100191, China +3

Remote Sensing Image Understanding Image Captioning Remote Sensing Image Change Captioning Representation Disentanglement Remote Sensing Change Detection

Ask PDF

Jul 3, 2026

LBTCap: A Lightweight Bilateral Transformer for Real-Time Remote Sensing Image Change Captioning

Remote sensing image change captioning (RSICC) generates natural-language descriptions of semantic changes between paired remote sensing images (RSIs), supporting applications such as urban planning, disaster response, and environmental monitoring. Although recent methods achieve strong captioning accuracy, most overlook computational efficiency and inference speed, which are essential for real-time deployment in practice. To this end, we propose LBTCap, a lightweight RSICC framework built on a bilateral Transformer that jointly models pre- and post-change features for efficient processing of paired RSIs. Specifically, we introduce a bilateral attention mechanism for paired inputs: the two temporal images are projected into separate queries and keys by the same query and key matrices shared across the two images, the value is formed from their concatenation, and the two resulting attention maps are combined by a learnable, structurally bilateral weighting instead of a fixed subtraction. This design keeps both temporal branches explicit while remaining compact, and, together with a truncated backbone and grouped-query attention, LBTCap uses only 39.99M parameters, of which the change-aware encoder accounts for just 2.78M. Extensive experiments on two public RSICC datasets show that LBTCap matches or closely approaches the accuracy of state-of-the-art methods while using far fewer parameters and running at markedly higher inference speed, with the benefit of the bilateral formulation most pronounced in the low-resource setting, demonstrating a favorable accuracy-efficiency trade-off for practical RSICC.

Licheng Zhang, Siew-Kei Lam, Naveed Akhtar

School of Computing and Information Systems, The University of Melbourne, Melbourne, 3010, Victoria, Australia · Nanyang Technological University, Singapore · The University of Melbourne, Melbourne, 3010, Victoria, Australia

Remote Sensing Image Understanding Image Captioning Remote Sensing Image Change Captioning Efficient ViTs Attention Mechanisms

Ask PDF

Jun 26, 2026

RSICCLLM: A Multimodal Large Language Model for Remote Sensing Image Change Captioning

Remote Sensing Image Change Captioning (RSICC) aims to describe changes between bi-temporal remote sensing images and holds significant research and application value. However, most existing methods rely on conventional deep learning architectures, and the limited model capacity constrains performance. Although large-model post-training techniques have achieved great success in general domains, their direct transfer to RSICC remains challenging due to data scarcity and the need for fine-grained change understanding. To address this, we propose RSICCLLM, the first post-training framework for large vision-language models in RSICC. Specifically, we design a data generation paradigm, release the instruction dataset RSICI, and establish a task-specific RSICC benchmark. We further introduce Difference-aware Supervised Fine-tuning to explicitly extract change representations and guide the model in perceiving and understanding temporal differences. In addition, we propose Dual-Negative Preference Optimization (DNPO), which employs two complementary negative-sample construction strategies to construct the preference dataset RSICP and further refine model performance. Extensive experiments validate the superior capability of RSICCLLM, which achieves outstanding results with only 7B parameters, surpassing models of substantially larger scales. The code and dataset will be made publicly available at https://github.com/keaill/RSICCLLM.

Yelin Wang, Zijia Song, Shuo Ye +6

State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · Harbin Institute of Technology, Shenzhen, China · Great Bay University, Dongguan, China +1

Remote Sensing Image Understanding VLM Adaptation Remote Sensing Image Change Captioning Multimodal Large Language Models Preference Optimization Remote Sensing Benchmarks

Ask PDF

Abstract

Remote Sensing Image Change Captioning

Explore similar work

HiSem: Hierarchical Semantic Disentangling for Remote Sensing Image Change Captioning

LBTCap: A Lightweight Bilateral Transformer for Real-Time Remote Sensing Image Change Captioning

RSICCLLM: A Multimodal Large Language Model for Remote Sensing Image Change Captioning