Global-Local Contextual Progressive Expansion Network for Martian Landslide Segmentation in Multimodal Remote Sensing Imagery

Leo Thomas Ramos; Sidike Paheding; Abel A. Reyes-Angulo; Rajaneesh A.; Sajinkumar K. S.; Angel D. Sappa; Thomas Oommen

cs.CVSep 11, 2026

Global-Local Contextual Progressive Expansion Network for Martian Landslide Segmentation in Multimodal Remote Sensing Imagery

Authors: Leo Thomas Ramos, Sidike Paheding, Abel A. Reyes-Angulo, Rajaneesh A., Sajinkumar K. S., Angel D. Sappa, Thomas Oommen

Organizations: Computer Vision Center, Universitat Aut`onoma de Barcelona, Barcelona, 08193, Spain · Department of Computer Science and Engineering, Fairfield University, Fairfield, CT 06824, USA · Michigan Technological University, Houghton, MI 49931, USA · Department of Geology, University of Kerala, Thiruvananthapuram 695581, Kerala, India ·

Abstract

Automated landslide segmentation on Mars is one of the important tasks for understanding its surface processes, and all will aid in future space exploration. However, it remains a relatively underexplored open challenge because landslide morphology is highly variable, foreground regions are often sparse or irregular, and orbital observations combine heterogeneous spectral and topographic cues. In this context, this work investigates the capability of deep learning to address Martian landslide segmentation through an extensive assessment of modern neural segmentation models. To the best of our knowledge, this is the first study to provide such a comprehensive exploration in this domain. We further propose TransCPLES, a U-shaped network that couples Contextual Progressive Layer Expansion feature extraction with Transformer-based contextual reasoning, enabling the model to capture local geomorphic patterns and broader spatial dependencies for more reliable landslide delineation. Experiments on MMLSv2, a seven-band multimodal Martian landslide dataset, show that TransCPLES achieves the best overall performance when evaluated on geographically distinct samples, with consistent delineation across different landslide extents, stable foreground discrimination, and a favorable balance between accuracy and computational cost compared with several state-of-the-art convolutional, attention-based, and Transformer-based segmentation models. With this work, we hope to provide a useful reference and encourage further research and development in deep learning for planetary remote sensing. Code will be available after publication.

Figures & tables

Fig. 1 : Geographic distribution of the standard and isolated sets comprising the MMLSv2 dataset.

Fig. 2 : Example scenes from the MMLSv2 dataset. Channel order: (1) Red, (2) Green, (3) Blue, (4) DEM, (5) Slope, (6) Thermal inertia, (7) Grayscale.

Split	# Images	Avg. FG (%)	Std. FG (%)	Min. FG (%)	Max. FG (%)
Train	465	35.41	25.64	0.02	99.52
Val	66	31.53	24.05	0.08	90.32
Test	133	33.82	25.05	0.10	90.67
Isolated test	276	21.83	17.08	0.01	71.95

TABLE I : Breakdown of the MMLSv2 dataset across its respective splits. The foreground proportion is reported as the percentage of pixels corresponding to landslide areas, including the mean (Avg.

{FG}

), standard deviation (Std.

{FG}

), and range defined by minimum and maximum values (Min.

{FG}

, Max.

{FG}

Fig. 3 : MarsLS-Net [ 22 ] architecture, including the overall framework with stacked PEN-attention blocks, the internal structure of the PEN-Attention module, the multi-head self-attention mechanism, and the ConvPE layer.

Fig. 4 : Proposed TransCPLES architecture, comprising an encoder–decoder framework with C-PLES blocks, and a Transformer-based bridge for global contextual reasoning.

Filters	Kernel size	Standard test set					Isolated test set				Params (M)
		IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
16	1	83.883 $\pm$ 0.594	70.800 $\pm$ 0.792	77.341 $\pm$ 0.579	0.089 $\pm$ 0.027	2.137 $\pm$ 0.000	86.003 $\pm$ 1.040	62.199 $\pm$ 0.852	74.101 $\pm$ 0.900	0.072 $\pm$ 0.000	0.075
	3	85.890 $\pm$ 0.198	74.364 $\pm$ 0.269	80.127 $\pm$ 0.068	0.074 $\pm$ 0.001	2.140 $\pm$ 0.000	86.161 $\pm$ 1.031	62.633 $\pm$ 1.857	74.397 $\pm$ 1.331	0.073 $\pm$ 0.000	0.109
	5	85.974 $\pm$ 0.343	74.463 $\pm$ 0.788	80.218 $\pm$ 0.507	0.073 $\pm$ 0.000	2.159 $\pm$ 0.024	86.839 $\pm$ 0.867	63.242 $\pm$ 2.110	75.040 $\pm$ 1.432	0.073 $\pm$ 0.000	0.176
32	1	83.972 $\pm$ 0.522	70.452 $\pm$ 0.643	77.212 $\pm$ 0.578	0.081 $\pm$ 0.000	2.212 $\pm$ 0.022	84.134 $\pm$ 1.037	59.566 $\pm$ 0.674	71.850 $\pm$ 0.843	0.081 $\pm$ 0.000	0.298
	3	85.935 $\pm$ 0.465	74.523 $\pm$ 0.491	80.229 $\pm$ 0.460	0.082 $\pm$ 0.002	2.192 $\pm$ 0.000	85.673 $\pm$ 1.652	62.387 $\pm$ 2.460	74.030 $\pm$ 2.053	0.081 $\pm$ 0.000	0.431

TABLE II : Ablation study of the number of convolutional filters and kernel size in MarsLS-Net.

PE terms	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
1	86.193 $\pm$ 0.299	74.603 $\pm$ 0.408	80.398 $\pm$ 0.354	0.115 $\pm$ 0.060	2.213 $\pm$ 0.018	86.495 $\pm$ 1.563	63.146 $\pm$ 1.720	74.821 $\pm$ 1.641	0.078 $\pm$ 0.000	0.869
2	86.317 $\pm$ 0.629	74.712 $\pm$ 1.215	80.515 $\pm$ 0.922	0.088 $\pm$ 0.000	2.293 $\pm$ 0.024	87.990 $\pm$ 0.428	64.877 $\pm$ 1.101	76.434 $\pm$ 0.695	0.088 $\pm$ 0.000	1.562
3	86.368 $\pm$ 0.576	74.798 $\pm$ 1.690	80.583 $\pm$ 1.133	0.158 $\pm$ 0.050	2.565 $\pm$ 0.011	87.489 $\pm$ 1.277	63.949 $\pm$ 2.103	75.719 $\pm$ 1.664	0.127 $\pm$ 0.000	2.549

TABLE III : Effect of the number of progressive expansion terms in MarsLS-Net.

PEN blocks	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
2	86.351 $\pm$ 0.283	74.641 $\pm$ 1.090	80.496 $\pm$ 0.649	0.058 $\pm$ 0.021	1.181 $\pm$ 0.002	87.737 $\pm$ 1.421	64.227 $\pm$ 1.381	75.982 $\pm$ 1.393	0.045 $\pm$ 0.000	0.785
3	86.589 $\pm$ 0.208	74.874 $\pm$ 0.638	80.732 $\pm$ 0.422	0.068 $\pm$ 0.000	1.744 $\pm$ 0.001	86.961 $\pm$ 1.707	63.413 $\pm$ 1.783	75.187 $\pm$ 1.741	0.067 $\pm$ 0.000	1.173
4	86.317 $\pm$ 0.629	74.712 $\pm$ 1.215	80.515 $\pm$ 0.922	0.088 $\pm$ 0.000	2.293 $\pm$ 0.024	87.990 $\pm$ 0.428	64.877 $\pm$ 1.101	76.434 $\pm$ 0.695	0.088 $\pm$ 0.000	1.562

TABLE IV : Effect of the number of stacked PEN-Attention blocks in MarsLS-Net.

MHSA heads	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
1	86.329 $\pm$ 0.074	74.618 $\pm$ 0.709	80.474 $\pm$ 0.389	0.083 $\pm$ 0.028	1.565 $\pm$ 0.042	87.869 $\pm$ 0.760	64.005 $\pm$ 0.947	75.937 $\pm$ 0.851	0.086 $\pm$ 0.000	1.562
2	86.317 $\pm$ 0.629	74.712 $\pm$ 1.215	80.515 $\pm$ 0.922	0.088 $\pm$ 0.000	2.293 $\pm$ 0.024	87.990 $\pm$ 0.428	64.877 $\pm$ 1.101	76.434 $\pm$ 0.695	0.088 $\pm$ 0.000	1.562
4	82.816 $\pm$ 0.781	66.609 $\pm$ 3.461	74.712 $\pm$ 2.121	0.145 $\pm$ 0.000	4.721 $\pm$ 2.616	86.289 $\pm$ 1.136	58.339 $\pm$ 2.128	72.314 $\pm$ 1.073	0.144 $\pm$ 0.000	1.562

TABLE V : Effect of the number of spatial self-attention heads in MarsLS-Net.

Variant	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
w/o convpe	76.384 $\pm$ 0.567	64.830 $\pm$ 0.795	70.607 $\pm$ 0.401	0.070 $\pm$ 0.000	2.144 $\pm$ 0.001	82.122 $\pm$ 0.580	55.986 $\pm$ 1.331	69.054 $\pm$ 0.743	0.069 $\pm$ 0.000	0.297
w/o pe	82.031 $\pm$ 0.526	67.885 $\pm$ 0.430	74.958 $\pm$ 0.451	0.078 $\pm$ 0.003	2.227 $\pm$ 0.001	85.156 $\pm$ 1.520	61.619 $\pm$ 2.633	73.387 $\pm$ 2.037	0.077 $\pm$ 0.000	0.869
w/o mhsa	84.456 $\pm$ 0.454	70.954 $\pm$ 0.139	77.705 $\pm$ 0.284	0.005 $\pm$ 0.001	0.312 $\pm$ 0.002	89.487 $\pm$ 0.103	61.734 $\pm$ 0.365	75.611 $\pm$ 0.227	0.004 $\pm$ 0.000	0.969
full	86.317 $\pm$ 0.629	74.712 $\pm$ 1.215	80.515 $\pm$ 0.922	0.088 $\pm$ 0.000	2.293 $\pm$ 0.024	87.990 $\pm$ 0.428	64.877 $\pm$ 1.101	76.434 $\pm$ 0.695	0.088 $\pm$ 0.000	1.562

TABLE VI : Component-wise ablation study of MarsLS-Net.

Band	Standard test set					Isolated test set
combination	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$
RGB	74.714 $\pm$ 1.667	54.144 $\pm$ 3.094	64.429 $\pm$ 2.378	0.094 $\pm$ 0.005	2.339 $\pm$ 0.002	75.245 $\pm$ 1.880	32.719 $\pm$ 3.099	53.982 $\pm$ 1.302	0.090 $\pm$ 0.000
RGB-DEM	72.790 $\pm$ 0.947	54.464 $\pm$ 1.959	63.627 $\pm$ 0.548	0.091 $\pm$ 0.000	2.343 $\pm$ 0.001	69.823 $\pm$ 5.487	37.860 $\pm$ 7.440	53.842 $\pm$ 5.189	0.090 $\pm$ 0.000
RGB-DEM-Slope	82.796 $\pm$ 0.983	68.020 $\pm$ 2.915	75.408 $\pm$ 1.870	0.092 $\pm$ 0.001	2.343 $\pm$ 0.003	87.167 $\pm$ 0.735	58.743 $\pm$ 6.702	72.955 $\pm$ 3.664	0.090 $\pm$ 0.000
RGB-DEM-Slope-Thermal	82.729 $\pm$ 0.358	67.366 $\pm$ 2.948	75.048 $\pm$ 1.653	0.091 $\pm$ 0.000	2.343 $\pm$ 0.001	87.539 $\pm$ 0.393	60.726 $\pm$ 4.820	74.133 $\pm$ 2.504	0.090 $\pm$ 0.000
RGB-DEM-Slope-Thermal-Gray	86.317 $\pm$ 0.629	74.712 $\pm$ 1.215	80.515 $\pm$ 0.922	0.093 $\pm$ 0.002	2.293 $\pm$ 0.024	87.990 $\pm$ 0.428	64.877 $\pm$ 1.101	76.434 $\pm$ 0.695	0.090 $\pm$ 0.000

TABLE VII : Impact of spectral-topographic band combinations on MarsLS-Net performance.

Fig. 5 : Visual inference results of MarsLS-Net under different input-band combinations from the MMLSv2 dataset. Prediction maps use green for true positives, red for false positives, blue for false negatives, and black for true negatives.

C-PLES heads	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
1	85.786 $\pm$ 0.780	72.652 $\pm$ 1.805	79.219 $\pm$ 1.293	0.018 $\pm$ 0.004	0.720 $\pm$ 0.002	89.033 $\pm$ 0.261	64.640 $\pm$ 1.009	76.837 $\pm$ 0.635	0.015 $\pm$ 0.000	14.624
2	86.621 $\pm$ 0.461	75.708 $\pm$ 1.089	81.165 $\pm$ 0.775	0.025 $\pm$ 0.003	1.046 $\pm$ 0.006	88.716 $\pm$ 0.088	66.645 $\pm$ 0.722	77.681 $\pm$ 0.405	0.024 $\pm$ 0.000	14.624
4	86.362 $\pm$ 0.387	74.806 $\pm$ 1.744	80.584 $\pm$ 1.557	0.041 $\pm$ 0.001	1.746 $\pm$ 0.006	88.806 $\pm$ 0.479	66.248 $\pm$ 0.454	77.527 $\pm$ 0.440	0.040 $\pm$ 0.000	14.624

TABLE VIII : Effect of the number of attention heads in the C-PLES skip-attention blocks of TransCPLES.

Bridge heads	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
1	86.517 $\pm$ 0.437	74.923 $\pm$ 0.946	80.720 $\pm$ 0.691	0.024 $\pm$ 0.066	1.042 $\pm$ 0.006	88.640 $\pm$ 0.302	66.384 $\pm$ 0.351	77.512 $\pm$ 0.287	0.025 $\pm$ 0.000	14.624
2	86.621 $\pm$ 0.461	75.708 $\pm$ 1.089	81.165 $\pm$ 0.775	0.025 $\pm$ 0.003	1.046 $\pm$ 0.006	88.716 $\pm$ 0.088	66.645 $\pm$ 0.722	77.681 $\pm$ 0.405	0.024 $\pm$ 0.000	14.624
4	86.255 $\pm$ 0.716	74.570 $\pm$ 1.310	80.413 $\pm$ 1.012	0.026 $\pm$ 0.000	1.049 $\pm$ 0.004	88.803 $\pm$ 0.067	66.023 $\pm$ 0.756	77.413 $\pm$ 0.408	0.026 $\pm$ 0.001	14.624

TABLE IX : Effect of the number of attention heads in the Transformer bridge of TransCPLES.

Bridge layers	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
1	86.147 $\pm$ 0.407	73.990 $\pm$ 0.493	80.068 $\pm$ 0.448	0.023 $\pm$ 0.048	1.023 $\pm$ 0.003	88.955 $\pm$ 0.120	66.111 $\pm$ 0.488	77.533 $\pm$ 0.304	0.024 $\pm$ 0.000	11.466
2	85.830 $\pm$ 0.496	73.740 $\pm$ 1.422	79.785 $\pm$ 0.951	0.026 $\pm$ 0.000	1.045 $\pm$ 0.003	88.902 $\pm$ 0.299	66.015 $\pm$ 0.717	77.458 $\pm$ 0.334	0.024 $\pm$ 0.000	12.519
4	86.621 $\pm$ 0.461	75.708 $\pm$ 1.089	81.165 $\pm$ 0.775	0.025 $\pm$ 0.003	1.046 $\pm$ 0.006	88.716 $\pm$ 0.088	66.645 $\pm$ 0.722	77.681 $\pm$ 0.405	0.024 $\pm$ 0.000	14.624

TABLE X : Effect of the number of Transformer layers in the bottleneck bridge of TransCPLES.

Variant	Standard test set					Isolated test set				Params (M)
	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	Params (M)
w/o pne	86.588 $\pm$ 0.308	73.981 $\pm$ 0.470	80.285 $\pm$ 0.384	0.045 $\pm$ 0.034	0.684 $\pm$ 0.004	88.576 $\pm$ 0.223	65.550 $\pm$ 0.355	77.063 $\pm$ 0.103	0.025 $\pm$ 0.000	10.118
w/o mhsa	85.579 $\pm$ 0.392	73.213 $\pm$ 0.686	79.396 $\pm$ 0.504	0.017 $\pm$ 0.002	0.473 $\pm$ 0.002	88.866 $\pm$ 0.164	65.150 $\pm$ 1.208	77.008 $\pm$ 0.547	0.015 $\pm$ 0.000	14.622
w/o attention	84.847 $\pm$ 0.172	71.855 $\pm$ 0.319	78.351 $\pm$ 0.245	0.006 $\pm$ 0.001	0.273 $\pm$ 0.002	88.719 $\pm$ 0.080	65.615 $\pm$ 0.099	77.167 $\pm$ 0.017	0.005 $\pm$ 0.000	8.720
full	86.621 $\pm$ 0.461	75.708 $\pm$ 1.089	81.165 $\pm$ 0.775	0.025 $\pm$ 0.003	1.046 $\pm$ 0.006	88.716 $\pm$ 0.088	66.645 $\pm$ 0.722	77.681 $\pm$ 0.405	0.024 $\pm$ 0.000	14.624

TABLE XI : Component-wise ablation study of TransCPLES.

Band	Standard test set					Isolated test set
combination	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$	T. time (h) $\downarrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	Latency (s) $\downarrow$
RGB	80.601 $\pm$ 1.483	65.617 $\pm$ 1.774	73.109 $\pm$ 1.624	0.029 $\pm$ 0.057	1.050 $\pm$ 0.007	76.555 $\pm$ 1.991	44.210 $\pm$ 2.114	60.383 $\pm$ 2.013	0.028 $\pm$ 0.001
RGB-DEM	84.187 $\pm$ 1.746	72.083 $\pm$ 2.914	78.135 $\pm$ 2.328	0.030 $\pm$ 0.001	1.045 $\pm$ 0.001	79.139 $\pm$ 2.702	53.316 $\pm$ 4.905	66.227 $\pm$ 3.801	0.027 $\pm$ 0.000
RGB-DEM-Slope	86.777 $\pm$ 0.264	73.595 $\pm$ 0.724	80.186 $\pm$ 0.493	0.029 $\pm$ 0.000	1.066 $\pm$ 0.018	88.903 $\pm$ 0.333	62.529 $\pm$ 1.276	75.716 $\pm$ 0.749	0.027 $\pm$ 0.001
RGB-DEM-Slope-Thermal	86.484 $\pm$ 0.257	74.941 $\pm$ 0.280	80.713 $\pm$ 0.262	0.029 $\pm$ 0.000	1.046 $\pm$ 0.004	88.942 $\pm$ 0.154	64.144 $\pm$ 0.222	76.543 $\pm$ 0.091	0.027 $\pm$ 0.000
RGB-DEM-Slope-Thermal-Gray	86.621 $\pm$ 0.461	75.708 $\pm$ 1.089	81.165 $\pm$ 0.775	0.028 $\pm$ 0.003	1.046 $\pm$ 0.006	88.716 $\pm$ 0.088	66.645 $\pm$ 0.722	77.681 $\pm$ 0.405	0.026 $\pm$ 0.001

TABLE XII : Impact of spectral-topographic band combinations on TransCPLES performance.

Fig. 6 : Visual inference results of TransCPLES under different input-band combinations from the MMLSv2 dataset. Prediction maps use green for true positives, red for false positives, blue for false negatives, and black for true negatives.

Model	Precision (%) $\uparrow$	Recall (%) $\uparrow$	F1-score (%) $\uparrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	T. time (h) $\downarrow$	Latency (s) $\downarrow$
U-Net [ 43 ]	87.637 $\pm$ 2.033	78.749 $\pm$ 5.571	82.836 $\pm$ 2.109	85.035 $\pm$ 0.862	70.738 $\pm$ 3.106	77.887 $\pm$ 1.981	0.276 $\pm$ 0.042	0.014 $\pm$ 0.018
ResUNet [ 44 ]	86.345 $\pm$ 1.329	85.115 $\pm$ 3.080	85.690 $\pm$ 1.080	86.518 $\pm$ 0.583	74.973 $\pm$ 1.644	80.746 $\pm$ 1.096	0.375 $\pm$ 0.064	0.006 $\pm$ 0.000
UNet++ [ 45 ]	87.019 $\pm$ 0.982	83.960 $\pm$ 0.815	85.457 $\pm$ 0.468	86.501 $\pm$ 0.458	74.609 $\pm$ 0.712	80.555 $\pm$ 0.576	0.411 $\pm$ 0.003	0.008 $\pm$ 0.000
SK-UNet [ 46 ]	88.469 $\pm$ 1.933	82.011 $\pm$ 3.698	85.055 $\pm$ 1.168	86.540 $\pm$ 0.439	74.008 $\pm$ 1.770	80.274 $\pm$ 1.093	0.542 $\pm$ 0.001	0.014 $\pm$ 0.002
DeepLabV3+ [ 47 ]	85.764 $\pm$ 1.531	83.961 $\pm$ 2.235	84.830 $\pm$ 0.835	85.822 $\pm$ 0.626	73.662 $\pm$ 1.262	79.742 $\pm$ 0.911	0.547 $\pm$ 0.099	0.008 $\pm$ 0.003
SegFormer [ 48 ]	83.627 $\pm$ 0.196	86.616 $\pm$ 0.750	85.095 $\pm$ 0.442	85.486 $\pm$ 0.329	74.058 $\pm$ 0.668	79.772 $\pm$ 0.498	2.717 $\pm$ 0.834	0.040 $\pm$ 0.000

TABLE XIII : Performance comparison on the standard test set.

Fig. 7 : Qualitative segmentation comparison on the standard test set. Prediction maps use green for true positives, red for false positives, blue for false negatives, and black for true negatives.

Model	Precision (%) $\uparrow$	Recall (%) $\uparrow$	F1-score (%) $\uparrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	IoU FG Drop $\downarrow$	mIoU Drop $\downarrow$
U-Net [ 43 ]	74.965 $\pm$ 4.558	80.288 $\pm$ 6.180	77.293 $\pm$ 0.396	87.513 $\pm$ 0.987	62.991 $\pm$ 0.526	75.252 $\pm$ 0.415	7.747 $\pm$ 2.746	2.635 $\pm$ 2.281
ResUNet [ 44 ]	70.589 $\pm$ 3.851	82.281 $\pm$ 6.067	75.781 $\pm$ 0.819	86.014 $\pm$ 0.969	61.011 $\pm$ 1.063	73.512 $\pm$ 0.626	13.962 $\pm$ 1.204	7.234 $\pm$ 1.258
UNet++ [ 45 ]	69.542 $\pm$ 1.982	84.078 $\pm$ 2.665	76.078 $\pm$ 0.384	85.860 $\pm$ 0.618	61.393 $\pm$ 0.499	73.626 $\pm$ 0.481	13.217 $\pm$ 0.261	6.929 $\pm$ 0.379
SK-UNet [ 46 ]	69.950 $\pm$ 1.965	80.709 $\pm$ 4.944	74.845 $\pm$ 1.083	85.648 $\pm$ 0.269	59.810 $\pm$ 1.376	72.729 $\pm$ 0.560	14.198 $\pm$ 0.848	7.545 $\pm$ 0.718
DeepLabV3+ [ 47 ]	70.946 $\pm$ 4.612	84.827 $\pm$ 2.384	77.159 $\pm$ 1.894	86.491 $\pm$ 1.900	62.837 $\pm$ 2.489	74.664 $\pm$ 2.191	10.825 $\pm$ 3.719	5.078 $\pm$ 3.047
SegFormer [ 48 ]	71.818 $\pm$ 1.342	77.555 $\pm$ 2.671	74.553 $\pm$ 1.232	86.096 $\pm$ 0.528	59.441 $\pm$ 1.558	72.768 $\pm$ 1.003	14.617 $\pm$ 2.068	7.004 $\pm$ 1.335

TABLE XIV : Model performance on the isolated test set under geographic shift. Performance drop is reported in percentage points as: Standard - Isolated .

Fig. 8 : Qualitative segmentation comparison on the isolated test set. Prediction maps use green for true positives, red for false positives, blue for false negatives, and black for true negatives.

			Latency (s)
Model	T. time (h)	Mean epoch (s)	Standard	Isolated	FLOPS (G)	MACs (G)	Params (M)
U-Net [ 43 ]	0.276 $\pm$ 0.042	5.860 $\pm$ 1.216	0.014 $\pm$ 0.018	0.012 $\pm$ 0.000	24.269	12.134	28.957
ResUNet [ 44 ]	0.375 $\pm$ 0.064	8.243 $\pm$ 2.102	0.006 $\pm$ 0.000	0.007 $\pm$ 0.000	5.462	2.731	32.534
UNet++ [ 45 ]	0.411 $\pm$ 0.003	8.863 $\pm$ 0.060	0.008 $\pm$ 0.000	0.008 $\pm$ 0.000	28.898	14.449	48.998
SK-UNet [ 46 ]	0.542 $\pm$ 0.001	12.304 $\pm$ 0.023	0.014 $\pm$ 0.002	0.012 $\pm$ 0.000	5.704	2.852	34.457
DeepLabV3+ [ 47 ]	0.547 $\pm$ 0.099	12.148 $\pm$ 3.384	0.008 $\pm$ 0.003	0.006 $\pm$ 0.000	4.711	2.355	26.690

TABLE XV : Computational efficiency comparison across segmentation models.

Panel of Figure 24 — Fig. 9 : Accuracy-efficiency trade-off across segmentation models. (a) Isolated mIoU versus inference latency on the isolated test set. (b) Isolated mIoU versus total training time. (c) Computational complexity in terms of FLOPs and trainable parameters.

Model	Standard AP FG (%) $\uparrow$	Isolated AP FG (%) $\uparrow$	AP FG Drop $\downarrow$
U-Net [ 43 ]	90.716 $\pm$ 0.285	78.645 $\pm$ 1.359	12.072 $\pm$ 1.530
ResUNet [ 44 ]	91.122 $\pm$ 1.919	76.825 $\pm$ 0.253	14.297 $\pm$ 1.827
UNet++ [ 45 ]	92.221 $\pm$ 0.174	77.968 $\pm$ 0.383	14.252 $\pm$ 0.545
SK-UNet [ 46 ]	91.977 $\pm$ 0.703	76.888 $\pm$ 0.876	15.089 $\pm$ 1.577
DeepLabV3+ [ 47 ]	91.312 $\pm$ 1.045	78.958 $\pm$ 2.755	12.354 $\pm$ 3.481
SegFormer [ 48 ]	91.612 $\pm$ 0.658	78.207 $\pm$ 2.754	13.405 $\pm$ 2.678

TABLE XVI : Average precision comparison for the foreground landslide class.

Panel of Figure 26 — Fig. 10 : Foreground precision-recall curves on the standard and isolated test sets.

Model	Standard test set				Isolated test set
	Precision (%) $\uparrow$	Recall (%) $\uparrow$	F1-score (%) $\uparrow$	IoU (%) $\uparrow$	Precision (%) $\uparrow$	Recall (%) $\uparrow$	F1-score (%) $\uparrow$	IoU (%) $\uparrow$
U-Net [ 43 ]	29.258 $\pm$ 1.718	41.091 $\pm$ 2.447	34.179 $\pm$ 2.018	19.264 $\pm$ 1.290	37.844 $\pm$ 0.417	46.998 $\pm$ 1.581	41.917 $\pm$ 0.602	24.435 $\pm$ 0.348
ResUNet [ 44 ]	34.374 $\pm$ 1.120	40.842 $\pm$ 2.101	37.323 $\pm$ 1.438	21.233 $\pm$ 0.960	37.229 $\pm$ 0.482	41.868 $\pm$ 2.707	39.388 $\pm$ 1.350	22.610 $\pm$ 0.855
UNet++ [ 45 ]	35.856 $\pm$ 1.570	39.446 $\pm$ 0.409	37.556 $\pm$ 0.988	21.265 $\pm$ 0.558	39.713 $\pm$ 1.012	42.225 $\pm$ 1.484	40.922 $\pm$ 0.998	23.435 $\pm$ 0.579
SK-UNet [ 46 ]	35.340 $\pm$ 0.410	38.547 $\pm$ 1.514	36.859 $\pm$ 0.522	20.820 $\pm$ 0.351	38.385 $\pm$ 1.196	41.359 $\pm$ 2.041	39.813 $\pm$ 1.569	22.844 $\pm$ 0.977
DeepLabV3+ [ 47 ]	31.729 $\pm$ 1.167	39.435 $\pm$ 0.253	35.160 $\pm$ 0.814	19.889 $\pm$ 0.469	37.445 $\pm$ 2.943	42.560 $\pm$ 1.796	39.829 $\pm$ 2.457	22.912 $\pm$ 1.633

TABLE XVII : Boundary delineation accuracy across evaluated models under a 1-pixel tolerance.

Panel of Figure 28 — Fig. 11 : Boundary delineation comparison across segmentation models, 1-pixel tolerance.

		Standard test set			Isolated test set
Model	Bin	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$	IoU BG (%) $\uparrow$	IoU FG (%) $\uparrow$	mIoU (%) $\uparrow$
U-Net [ 43 ]	small	94.491 $\pm$ 0.639	35.580 $\pm$ 2.507	65.036 $\pm$ 1.529	91.824 $\pm$ 1.655	25.577 $\pm$ 1.463	58.701 $\pm$ 1.443
	medium	87.110 $\pm$ 0.222	62.390 $\pm$ 2.497	74.750 $\pm$ 1.332	86.690 $\pm$ 0.993	63.201 $\pm$ 0.582	74.945 $\pm$ 0.222
	large	71.266 $\pm$ 2.643	77.174 $\pm$ 3.864	74.220 $\pm$ 3.253	78.254 $\pm$ 0.651	75.710 $\pm$ 2.419	76.982 $\pm$ 1.522
ResUNet [ 44 ]	small	93.843 $\pm$ 0.626	32.848 $\pm$ 3.860	63.345 $\pm$ 2.201	89.980 $\pm$ 1.713	21.999 $\pm$ 1.042	55.990 $\pm$ 1.268
	medium	87.348 $\pm$ 0.446	64.491 $\pm$ 1.738	75.919 $\pm$ 0.983	85.264 $\pm$ 0.883	61.520 $\pm$ 1.453	73.392 $\pm$ 0.698

TABLE XVIII : Performance stratified by foreground-area ratio on the standard and isolated test sets. Small, medium, and large correspond to foreground ratios of

[0,10)

[10,40)

%, and

[40,100]

%, respectively.

Panel of Figure 30 — Fig. 12 : Qualitative segmentation inference stratified by foreground-area ratio on the isolated test set. Samples are grouped according to the foreground occupancy of the ground-truth mask. Prediction maps use green for true positives, red for false positives, blue for false negatives, and black for true negatives.

Multi-Angular Remote Sensing

See all 8 papers

DualSwinFusionSeg: Multimodal Martian Landslide Segmentation via Dual Swin Transformer with Multi-Scale Fusion and UNet++
Mar 14, 2026Shahriar Kabir, Abdullah Muhammed Amimul Ehsan, Istiak Ahmmed Rifti +1Multimodal Semantic Segmentation Image Segmentation
Bangladesh University of Engineering and Technology, Bangladesh · University of California, Riverside, USA
MANTLE: A Framework for Adaptive In-Situ Planetary Perception Using a Modular Uplink Principle
Aug 28, 2026Pranav Durai, Gary DoranSpace Robotics Remote Sensing Image Segmentation
Stanford Center for Innovation in In Vivo Imaging Stanford University School of Medicine 318 Campus Drive, Room S040, Stanford, CA 94305, USA · Jet Propulsion Laboratory California Institute of Technology 4800 Oak Grove Drive, Pasadena, CA 91109, USA
A Benchmark Study of Segmentation Models and Adaptation Strategies for Landslide Detection from Satellite Imagery
Apr 17, 2026Md Kowsher, Weiwei Zhan, Chen ChenRemote Sensing Image Segmentation Vision Foundation Model Adaptation
University of Central Florida

Mar 14, 2026J/K move · Enter open · S save

DualSwinFusionSeg: Multimodal Martian Landslide Segmentation via Dual Swin Transformer with Multi-Scale Fusion and UNet++

Shahriar Kabir, Abdullah Muhammed Amimul Ehsan, Istiak Ahmmed Rifti, Md Kaykobad Reza

Bangladesh University of Engineering and Technology, Bangladesh · University of California, Riverside, USA

Multimodal Semantic Segmentation Image Segmentation Remote Sensing Image Segmentation Multi-Scale Feature Fusion

Open paper AskPDF arXiv

Automated segmentation of Martian landslides, particularly in tectonically active regions such as Valles Marineris,is important for planetary geology, hazard assessment, and future robotic exploration. However, detecting landslides from planetary imagery is challenging due to the heterogeneous nature of available sensing modalities and the limited number of labeled samples. Each observation combines RGB imagery with geophysical measurements such as digital elevation models, slope maps, thermal inertia, and contextual grayscale imagery, which differ significantly in resolution and statistical properties. To address these challenges, we propose DualSwinFusionSeg, a multimodal segmentation architecture that separates modality-specific feature extraction and performs multi-scale cross-modal fusion. The model employs two parallel Swin Transformer V2 encoders to independently process RGB and auxiliary geophysical inputs, producing hierarchical feature representations. Corresponding features from the two streams are fused at multiple scales and decoded using a UNet++ decoder with dense nested skip connections to preserve fine boundary details. Extensive ablation studies evaluate modality contributions, loss functions, decoder architectures, and fusion strategies. Experiments on the MMLSv2 dataset from the PBVS 2026 Mars-LS Challenge show that modality-specific encoders and simple concatenation-based fusion improve segmentation accuracy under limited training data. The final model achieves 0.867 mIoU and 0.905 F1 on the development benchmark and 0.783 mIoU on the held-out test set, demonstrating strong performance for multimodal planetary surface segmentation.

Abstract

Multi-Angular Remote Sensing

Explore similar work

DualSwinFusionSeg: Multimodal Martian Landslide Segmentation via Dual Swin Transformer with Multi-Scale Fusion and UNet++

MANTLE: A Framework for Adaptive In-Situ Planetary Perception Using a Modular Uplink Principle

A Benchmark Study of Segmentation Models and Adaptation Strategies for Landslide Detection from Satellite Imagery