Beyond Quadratic: Linear-Time Change Detection with RWKV
Authors: Zhenyu Yang, Gensheng Pei, Tao Chen, Xia Yuan, Haofeng Zhang, Xiangbo Shu, Yazhou Yao
Organizations: School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China · State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery, China · Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, Korea
Existing paradigms for remote sensing change detection are caught in a trade-off: CNNs excel at efficiency but lack global context, while Transformers capture long-range dependencies at a prohibitive computational cost. This paper introduces ChangeRWKV, a new architecture that reconciles this conflict. By building upon the Receptance Weighted Key Value (RWKV) framework, our ChangeRWKV uniquely combines the parallelizable training of Transformers with the linear-time inference of RNNs. Our approach core features two key innovations: a hierarchical RWKV encoder that builds multi-resolution feature representation, and a novel Spatial-Temporal Fusion Module (STFM) engineered to resolve spatial misalignments across scales while distilling fine-grained temporal discrepancies. ChangeRWKV not only achieves state-of-the-art performance on the LEVIR-CD benchmark, with an 85.46% IoU and 92.16% F1 score, but does so while drastically reducing parameters and FLOPs compared to previous leading methods. This work demonstrates a new, efficient, and powerful paradigm for operational-scale change detection. Our code and model are publicly available.
Figures & tables
Figure 1: Efficiency versus accuracy on the LEVIR-CD benchmark. The proposed ChangeRWKV family establishes a new state-of-the-art frontier, delivering superior IoU scores while demanding significantly fewer computational resources (FLOPs) and parameters than existing methods. Our tiny model, for instance, achieves a competitive 84.92% IoU with only 4.7M Params and 9.40G FLOPs.
Figure 2: Overall architecture of ChangeRWKV. The model consists of three main components: ( a) a Hierarchical RWKV Encoder that extracts multi-scale features, (b) a Spatial-Temporal Fusion Module (STFM) that integrates spatial and temporal cues, and (c) a lightweight Decoder that generates the change mask. The STFM is further decomposed into (b1) a Spatial Fusion Module (SFM) for multi-scale spatial alignment and (b2) a Temporal Fusion Module (TFM) for bi-temporal interaction.
Method
Params (M)
FLOPs (G)
LEVIR-CD
WHU-CD
IoU
F1
P
R
IoU
F1
P
R
FC-Siam-Diff ( 2018 )
4.38
1.35
75.92
86.31
89.53
83.31
41.66
58.81
47.33
77.66
FC-Siam-Conc ( 2018 )
4.99
1.55
71.96
83.69
91.99
76.77
49.95
66.63
60.88
73.58
SNUNet ( 2021 )
12.03
27.44
78.83
88.16
89.18
87.17
71.67
83.50
85.60
81.49
BIT ( 2021 )
3.55
4.35
80.68
89.31
89.24
89.37
72.39
83.98
86.64
81.48
ICIFNet ( 2022 )
23.82
25.36
81.75
89.96
91.32
88.64
79.24
88.32
92.98
85.56
Table 1: Quantitative comparison on the LEVIR-CD ( Chen and Shi 2020 ) and WHU-CD ( Hong et al. 2023 ) test sets. All metrics are reported as a percentage (%), with the highest value highlighted in bold , and the second highest underlined .
Method
Params (M)
FLOPs (G)
IoU
F1
FC-Siam-Diff ( 2018 )
4.38
1.35
57.44
72.97
FC-Siam-Conc ( 2018 )
4.99
1.55
58.07
73.48
STANet ( 2020 )
16.93
6.58
65.66
79.31
SNUNet ( 2021 )
12.03
27.44
67.11
80.32
BIT ( 2021 )
3.55
4.35
70.64
82.80
TransUNetCD ( 2022 )
28.37
244.54
71.86
83.63
Table 2: Quantitative comparison on the LEVIR-CD+ ( Shen et al. 2021 ) test set. All metrics are reported as a percentage (%). Best results are highlighted in bold .
Method
Params (M)
FLOPs (G)
IoU
F1
FC-Siam-Diff ( 2018 )
4.38
1.35
86.07
92.52
FC-Siam-Conc ( 2018 )
4.99
1.55
93.39
96.58
BIT ( 2021 )
3.55
4.35
93.48
96.63
MFPNet ( 2021 )
60.06
107.73
97.09
98.52
MSCANet ( 2022 )
16.59
14.68
94.43
97.14
FCS ( 2022 )
3.03
9.59
94.77
97.31
Table 3: Quantitative comparison on the SAR-CD ( Alatalo et al. 2023 ) test set. All metrics are reported as a percentage (%). Best results are highlighted in bold .
Figure 3: Qualitative results on the LEVIR-CD test set. Predicted outputs are color-coded as follows: white for true positives (TP), black for true negatives (TN), green for false positives (FP), and red for false negatives (FN).
Figure 4: Qualitative results on the SAR-CD test set. Predicted outputs are color-coded as follows: white for true positives (TP), black for true negatives (TN), green for false positives (FP), and red for false negatives (FN).
Experiment
Variant #
Method
IoU
F1
Component
Encoder Architectures
❶
VisionRWKV
81.86
90.03
❷
+ ViT-Adapter
83.06
90.75
-
ours
85.46
92.16
Channel Mixing Module
❸
w/o
84.43
91.56
❹
RWKV-like
85.38
92.11
Table 4: Ablation study on different architectural components. All metrics are reported as a percentage (%).
Figure 5: Computational Scalability Analysis. ChangeRWKV demonstrates near-linear growth in (a) FLOPs, (b) training memory, and (c) inference memory, significantly outperforming Transformer-based models at high resolutions.