Improved MambdaBDA Framework for Robust Building Damage Assessment Across Disaster Domains

Alp Eren Gençoğlu; Hazım Kemal Ekenel

cs.CVMar 1, 2026

Improved MambdaBDA Framework for Robust Building Damage Assessment Across Disaster Domains

Authors: Alp Eren Gençoğlu, Hazım Kemal Ekenel

Organizations: Department of Computer Engineering, Istanbul Technical University, Türkiye · Division of Engineering, New York University Abu Dhabi, UAE

Abstract

Reliable post-disaster building damage assessment (BDA) from satellite imagery is hindered by severe class imbalance, background clutter, and domain shift across disaster types and geographies. In this work, we address these problems and explore ways to improve the MambaBDA, the BDA network of ChangeMamba architecture, one of the most successful BDA models. The approach enhances the MambaBDA with three modular components: (i) Focal Loss to mitigate class imbalance damage classification, (ii) lightweight Attention Gates to suppress irrelevant context, and (iii) a compact Alignment Module to spatially warp pre-event features toward post-event content before decoding. We experiment on multiple satellite imagery datasets, including xBD, Pakistan Flooding, Turkey Earthquake, and Ida Hurricane, and conduct in-domain and crossdataset tests. The proposed modular enhancements yield consistent improvements over the baseline model, with 0.8% to 5% performance gains in-domain, and up to 27% on unseen disasters. This indicates that the proposed enhancements are especially beneficial for the generalization capability of the system.

Figures & tables

Figure 1: Network structure of MambaBDA, adapted from [ 5 ] .

Damage Level	Name	Pixel Value
L1	No Damage	1
L2	Minor Damage	2
L3	Major Damage	3
L4	Destroyed	4

Table 1: Damage Levels

Figure 2: Sample data from xBD consisting of image pairs and mask pairs.

	Damage Level	Training	Valid	Test
Number of Images	L1	4624	569	570
	L2	1785	277	244
	L3	1797	320	294
	L4	2041	332	288
	Total	9168	933	933
Pixel Ratio	L1	76.35 %	73.29 %	76.67 %

Table 2: xBD Dataset Statistics.

	Damage Level	Training	Valid	Test
Number of Images	L1	800	177	174
	L2	317	70	67
	L3	155	36	42
	L4	510	106	119
	Total	2478	531	531
Pixel Ratio	L1	79.58 %	80.87 %	79.86 %

Table 3: Pakistan Flooding Dataset Statistics.

	Damage Level	Training	Valid	Test
Number of Images	L1	489	154	166
	L2	80	22	19
	L3	48	13	11
	L4	102	33	30
	Total	566	189	189
Pixel Ratio	L1	95.35 %	96.36 %	97.04 %

Table 4: Turkey Earthquake Dataset Statistics.

	Damage Level	Test
Number of Images	L1	438
	L2	168
	L3	79
	L4	35
	Total	528
Pixel Ratio	L1	88.67 %

Table 5: Hurricane Ida Dataset Statistics.

Model / Module	Params (M)	GFLOPs
Baseline	87.76	195.43
FOCAL	+ 0	+ 0
AGBD	+ 0.10	+ 0.71
AGB	+ 0.05	+ 0.36
ALIGN	+ 0.63	+ 0.65

Table 6: Model Complexity Comparison

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	84.86	75.23	78.12	95.23	57.36	73.02	86.66
AGBD	84.48	74.98	77.83	95.13	55.55	75.17	86.74
AGB	85.42	76.29	79.03	94.75	59.32	74.71	85.92
ALIGN	85.14	75.76	78.57	95.45	57.27	75.19	86.48
FOCAL	84.84	75.77	78.49	94.56	57.97	74.30	86.90
FOCAL + ALIGN	84.68	76.12	78.69	95.38	58.41	75.10	85.99

Table 7: In-Domain Evaluation Results on xBD Dataset (Group 1)

Figure 3: Example attention gate outputs. Heatmaps are overlayed on ground truth masks. Attention increases from blue to green to red. The top right heatmap belongs to AGB, the AGs in building decoder, and the bottom right heatmap belongs to AGD, the AGs in damage decoder. AGB mostly focuses on building borders improving performance, while AGD mostly focuses on spaces between the buildings, which does not increase the scores.

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	87.67	77.44	80.51	96.66	74.03	69.24	74.87
AGBD	87.81	76.99	80.24	96.59	73.86	67.27	75.77
AGB	87.75	76.46	79.85	97.04	78.07	63.03	74.98
ALIGN	87.58	78.01	80.88	96.49	74.20	68.75	77.58
FOCAL	88.15	77.63	80.78	96.71	74.86	69.02	74.95
FOCAL + ALIGN	88.43	77.19	80.56	96.77	73.64	66.66	77.50

Table 8: In-Domain Evaluation Results on Pakistan Flooding Dataset (Group 2)

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	90.69	29.97	48.18	98.43	20.13	17.06	66.63
AGBD	90.75	35.00	51.72	98.96	22.44	22.09	69.63
AGB	91.31	36.86	53.20	98.94	27.14	21.27	68.76
ALIGN	91.28	29.06	47.73	98.98	15.75	20.05	70.61
FOCAL	91.19	30.33	48.59	98.92	20.29	17.17	70.28
FOCAL + ALIGN	91.22	29.51	48.02	98.92	15.14	22.29	68.77

Table 9: In-Domain Evaluation Results on Turkey Earthquake Dataset

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	60.27	16.40	29.56	91.30	5.42	41.44	41.01
AGBD	63.39	24.36	36.07	91.46	9.41	44.74	40.55
AGB	63.93	34.94	43.63	93.67	17.32	45.60	41.41
ALIGN	53.75	38.76	43.26	90.83	23.25	47.16	35.76
FOCAL	61.56	43.00	48.57	91.76	33.42	44.25	33.78
FOCAL + ALIGN	57.35	37.15	43.21	85.34	23.79	44.43	31.84

Table 10: Cross-Dataset Evaluation of xBD-Trained Models (Group 1) on Pakistan Flooding Dataset

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	88.95	19.98	40.67	90.20	21.54	23.75	9.95
AGBD	86.27	14.85	36.28	93.08	24.81	5.74	22.64
AGB	88.02	24.79	43.76	89.27	29.32	20.11	15.07
ALIGN	89.18	24.99	44.24	92.24	31.58	35.64	11.17
FOCAL	88.90	16.37	38.13	88.87	26.96	51.75	5.66
FOCAL + ALIGN	87.84	21.37	41.31	91.77	34.24	37.44	8.31

Table 11: Cross-Dataset Evaluation of xBD-Trained Models (Group 1) on Hurricane Ida Dataset

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	84.04	0.02	25.23	97.08	0.00	2.52	18.32
AGBD	84.08	0.00	25.22	97.89	0.00	1.09	18.50
AGB	84.36	0.72	25.81	97.76	0.20	2.23	28.40
ALIGN	82.95	0.16	24.99	96.43	0.04	1.63	17.73
FOCAL	84.63	0.08	25.45	96.83	0.02	1.22	25.61
FOCAL + ALIGN	82.90	0.00	24.87	95.10	0.00	11.25	15.36

Table 12: Cross-Dataset Evaluation of xBD-Trained Models (Group 1) on Turkey Earthquake Dataset

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	40.13	0.51	12.39	86.58	0.13	12.63	12.08
AGBD	33.59	12.22	18.63	86.37	4.75	34.17	13.18
AGB	31.64	5.95	13.65	77.46	1.79	23.24	17.36
ALIGN	14.10	2.83	6.21	86.78	0.80	16.27	10.78
FOCAL	33.64	6.54	14.67	83.65	2.20	27.44	9.19
FOCAL + ALIGN	48.13	6.03	18.66	80.55	1.83	24.91	15.87

Table 13: Cross-Dataset Evaluation of Pakistan Flooding-Trained Models (Group 2) on xBD Dataset

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	47.60	0.36	14.53	93.35	0.11	0.62	2.88
AGBD	46.21	2.62	15.69	91.63	1.50	16.95	1.26
AGB	39.55	3.45	14.28	87.16	1.53	11.61	2.46
ALIGN	29.79	0.00	8.94	93.70	0.79	0.00	4.56
FOCAL	42.89	3.59	15.38	90.46	1.85	7.74	2.30
FOCAL + ALIGN	51.78	0.48	15.87	89.61	0.13	19.49	3.34

Table 14: Cross-Dataset Evaluation of Pakistan Flooding-Trained Models (Group 2) on Hurricane Ida Dataset

Model	$F_{1}^{\text{loc}}$	$F_{1}^{\text{clf}}$	$F_{1}^{\text{oa}}$	$F_{1}^{\text{L1}}$	$F_{1}^{\text{L2}}$	$F_{1}^{\text{L3}}$	$F_{1}^{\text{L4}}$
Baseline	64.93	0.00	19.48	96.95	0.57	0.00	4.47
AGBD	55.88	0.00	16.77	96.99	0.11	0.00	2.28
AGB	54.74	0.24	16.59	92.12	1.43	0.06	7.75
ALIGN	52.87	0.00	15.86	97.26	0.77	0.00	1.96
FOCAL	62.27	1.18	19.50	94.86	1.69	0.38	6.67
FOCAL + ALIGN	64.17	0.46	19.57	92.21	0.20	0.28	2.94

Table 15: Cross-Dataset Evaluation of Pakistan Flooding-Trained Models (Group 2) on Turkey Earthquake Dataset

Disaster Damage Assessment

Disaster damage assessment uses imagery and geospatial data to locate, classify, and quantify damage for emergency response and recovery.

See all 48 papers

Robust Building Damage Detection in Cross-Disaster Settings Using Domain Adaptation
Mar 16, 2026Asmae Mouradi, Shruti KshirsagarDisaster Damage Assessment Building Damage Assessment
School of Computing, Wichita State University, Wichita, Kansas, USA
Revisiting the Task of Building Damage Assessment by Satellite Imaging
Oct 9, 2026Anthony Cioppa, Virgile Blua, Sébastien Piérard +2Building Damage Assessment Disaster Damage Assessment
HASTE: A Platform for Rapid Post-Disaster Building Damage Assessment
Jul 13, 2026Caleb Robinson, Anthony Ortiz, Simone Fobi Nsutezo +8Remote Sensing Image Segmentation Disaster Damage Assessment
Microsoft AI for Good Research Lab

Mar 16, 2026J/K move · Enter open · S save

Robust Building Damage Detection in Cross-Disaster Settings Using Domain Adaptation

Asmae Mouradi, Shruti Kshirsagar

School of Computing, Wichita State University, Wichita, Kansas, USA

Disaster Damage Assessment Building Damage Assessment Domain Adaptation

Open paper AskPDF arXiv

Rapid structural damage assessment from remote sensing imagery is essential for timely disaster response. Within human-machine systems (HMS) for disaster management, automated damage detection provides decision-makers with actionable situational awareness. However, models trained on multi-disaster benchmarks often underperform in unseen geographic regions due to domain shift - a distributional mismatch between training and deployment data that undermines human trust in automated assessments. We explore a two-stage ensemble approach using supervised domain adaptation (SDA) for building damage classification across four severity classes. The pipeline adapts the xView2 first-place method to the Ida-BD dataset using SDA and systematically investigates the effect of individual augmentation components on classification performance. Comprehensive ablation experiments on the unseen Ida-BD test split demonstrate that SDA is indispensable: removing it causes damage detection to fail entirely. Our pipeline achieves the most robust performance using SDA with unsharp-enhanced RGB input, attaining a Macro-F1 of 0.5552. These results underscore the critical role of domain adaptation in building trustworthy automated damage assessment modules for HMS-integrated disaster response.

Abstract

Disaster Damage Assessment

Explore similar work

Robust Building Damage Detection in Cross-Disaster Settings Using Domain Adaptation

Revisiting the Task of Building Damage Assessment by Satellite Imaging

HASTE: A Platform for Rapid Post-Disaster Building Damage Assessment