nnMNet: Baseline for Martian Terrain Semantic Segmentation

Ming-Han Lee; Chi-Yeh Chen

cs.CVAug 30, 2026

nnMNet: Baseline for Martian Terrain Semantic Segmentation

Authors: Ming-Han Lee, Chi-Yeh Chen

Organizations: Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan 701, Taiwan

Abstract

Semantic segmentation is a crucial task for understanding Mars, the most Earth-like planet in our solar system. However, it is challenging because the Martian surface is highly unstructured and complex, making accurate pixel-level prediction and fine-grained annotation difficult. Recent advancements in deep learning have introduced numerous methods and datasets to address these challenges. Nevertheless, the field lacks a robust, publicly available, and reproducible baseline, as well as a unified benchmark to facilitate fair evaluations. In this work, we present nnMNet, a new baseline model designed for Martian terrain semantic segmentation. Building upon nnWNet, we integrate linear attention to better capture global context and employ lightweight convolutions to reduce computational overhead. To bridge the gap between local and global representations, we introduce the Spatially-Aware Fusion Block (SAFB), which augments and combines features with diverse characteristics. Furthermore, we establish a new benchmark by curating and standardizing three high-quality datasets for thorough evaluation. nnMNet achieves new state-of-the-art 86.61%, 83.25%, and 88.24% mIoU on SynMars-TW, SynMars-Air, and MarsScapes, respectively. Our code, models, and datasets are publicly available at https://github.com/dereklee0310/nnMNet.

Figures & tables

Fig. 1: Performance comparison on SynMars-TW. nn M Net achieves a new state-of-the-art 86.61% mIoU with low parameters and moderate FLOPs.

Dataset	Source	Type	Annotation Method	Scale	Classes
AI4Mars [ 59 ]	MER, MSL	Real	Crowdsourcing	16.39k	4
MER-Seg [ 60 ]	MER	Real	Average Labeling	1.06k	9
MSL-Seg [ 60 ]	MSL	Real	Average Labeling	4.13k	9
$\text{S}^{5}$ Mars [ 61 ]	MSL	Real	Sparse Labeling	6k	9
MarsScapes [ 62 , 35 ]	MSL	Real	Semi-automatic	20.8k	9
SynMars [ 5 ]	TW-1	Synthetic	VisionBlender	60k	1

TABLE I: Publicly Available Martian Terrain Semantic Segmentation Datasets

Fig. 2: Overview of nn M Net. OPE denotes overlapping patch embedding, LAB denotes Linear Attention Block, IRB denotes Inverted Residual Block, and SAFB denotes Spatially-Aware Fusion Block.

Panel of Figure 4 — Fig. 3: Heatmap visualization of the coarse and fine networks on SynMars-TW. They play complementary roles in identifying foreground objects.

Fig. 4: Comparison of the PoolFormer block (left) and Linear Attention Block (LAB) (right).

Fig. 5: Comparison of positional encodings on self-attention (left) and linear attention (right). RPE does not apply to linear attention.

Fig. 6: Illustration of the proposed Spatially-Aware Fusion Block.

Class	IoU (%)
Class	SynMars-TW	SynMars-Air	MarsScapes	MSL-Seg	$\text{S}^{5}$ Mars
Sky	98.43	99.54	83.06	—	83.83
Ridge	93.49	81.86	73.59	—	82.38
Soil / Martian Soil	98.79	99.11	86.89	17.24	30.39
Sand / Sands	77.73	65.33	73.91	41.00	37.69
Bedrock	68.53	70.18	75.13	59.11	57.87

TABLE II: nnUNet Results on Martian Terrain Segmentation Datasets

Panel of Figure 9 — Fig. 7: Examples from MSL-Seg (left) and S5Mars (right). The top row shows the input images and the bottom row shows the ground-truth labels. Red boxes highlight regions with ambiguous boundaries or missing labels.

Method	Big rock	Small rock	Gravel	Bedrock	Ridge	Sand	Soil	Sky	mIoU (%)	Params (M)	FLOPs (G)	FPS
BiseNetV2 [ 90 ]	83.58	22.09	0.01	84.43	95.95	68.02	96.09	99.41	68.70	3.35	12.31	501.13
DeepLabv3+ [ 21 ]	88.80	43.98	1.30	89.46	96.95	80.25	97.33	99.52	74.70	12.32	54.27	192.78
SegFormer-B0 [ 12 ]	89.74	42.86	1.41	91.06	95.56	84.43	97.65	99.30	75.25	3.75	8.50	302.28
SegNeXt-T [ 23 ]	89.24	31.43	0.23	89.41	95.88	83.83	97.16	99.36	73.32	4.26	6.45	221.09
Light4Mars-B [ 6 ]	89.81	44.37	0.09	91.70	96.65	84.78	97.84	99.47	76.70	2.56	9.42	160.51
nnUNet [ 15 ]	94.31	78.10	56.61	76.44	94.77	81.73	99.08	99.36	85.05	65.22	60.20	274.69

TABLE III: Comparison with state-of-the-art methods on SynMars-TW

Method	Big rock	Small rock	Gravel	Bedrock	Ridge	Sand	Soil	Sky	mIoU (%)	Params (M)	FLOPs (G)	FPS
BiseNetV2 [ 90 ]	84.74	29.14	0.00	90.00	98.27	96.63	98.66	99.77	74.65	3.35	12.31	501.13
DeepLabv3+ [ 21 ]	87.86	43.31	0.34	92.55	98.71	95.54	99.03	99.81	77.15	12.32	54.27	192.78
SegFormer-B0 [ 12 ]	90.56	52.67	1.09	94.33	98.83	98.24	99.26	99.75	79.34	3.75	8.50	302.28
SegNeXt-T [ 23 ]	91.36	43.62	0.07	94.02	98.97	98.21	99.10	99.83	78.15	4.26	6.45	221.09
Light4Mars-B [ 6 ]	89.98	49.81	5.02	94.38	98.94	97.95	99.26	99.83	79.40	2.56	9.42	160.51
nnUNet [ 15 ]	89.40	77.94	43.04	77.62	87.58	74.39	99.31	98.01	80.91	65.22	60.20	274.69

TABLE IV: Comparison with state-of-the-art methods on SynMars-Air

Method	Soil	Bedrock	Gravel	Sand	Big Rock	Ridge	Sky	Rover	Unlabeled	mIoU (%)	Params (M)	FLOPs (G)	FPS
BiseNetV2 [ 90 ]	89.30	84.20	74.16	81.16	63.43	88.68	95.47	77.78	79.22	81.49	3.35	6.16	623.24
DeepLabv3+ [ 21 ]	85.94	80.52	66.04	78.15	57.36	83.70	91.68	63.75	72.03	75.47	12.32	27.14	300.22
SegFormer-B0 [ 12 ]	91.12	86.59	78.25	84.16	67.50	92.46	96.61	88.64	84.25	85.51	3.72	3.70	395.43
SegNeXt-T [ 23 ]	91.71	87.21	79.81	84.18	67.98	93.00	96.92	88.11	83.87	85.87	4.26	3.23	252.14
Light4Mars-B [ 6 ]	90.78	86.59	77.11	83.64	66.24	91.35	95.97	83.68	81.96	84.15	2.56	4.82	282.06
nnUNet [ 15 ]	93.28	87.45	85.24	84.67	83.19	80.73	90.13	97.85	80.00	86.95	47.64	29.84	387.84

TABLE V: Comparison with state-of-the-art methods on MarsScapes

Method	mIoU (%)	Params (M)	FLOPs (G)
nn W Net [ 13 ]	81.17 ( $\uparrow$ 0.00)	7.05	42.54
Pooling $\rightarrow$ RALA [ 81 ]	81.62 ( $\uparrow$ 0.45)	7.21	43.91
+ RoPE [ 87 ]	82.05 ( $\uparrow$ 0.43)	7.21	43.91
+ APE [ 43 ]	81.64 ( $\downarrow$ 0.41)	9.18	43.91
+ CPE [ 68 , 84 ]	81.79 ( $\downarrow$ 0.26)	7.22	44.00
+ LePE [ 85 ]	81.78 ( $\downarrow$ 0.27)	7.22	44.00

TABLE VI: Ablation of nn M Net components

token mixer	mIoU (%)	Params (M)	FLOPs (G)
RALA [ 81 ]	82.55	5.71	39.81
Efficient Self-Attention [ 12 ]	81.99	6.47	44.11
Average Pooling [ 75 ]	82.13	5.55	38.46

TABLE VII: Ablation of the token mixer within LAB

kernel size	mIoU (%)	Params (M)	FLOPs (G)
3 $\times$ 3	82.55	5.71	39.81
5 $\times$ 5	82.36	5.77	40.56
7 $\times$ 7	82.33	5.85	41.69

TABLE VIII: Ablation of IRB kernel size

reduction ratio	mIoU (%)	Params (M)	FLOPs (G)
1	82.57	6.16	42.57
0.5	82.54	5.86	40.73
0.25	82.55	5.71	39.81
0.125	82.38	5.64	39.95
w/o SAFB	82.43	5.56	37.71

TABLE IX: Ablation of SAFB Reduction Ratio

Model	Iterations	mIoU (%)	Params (M)	FLOPs (G)	FPS
nn W Net	50K	80.02	7.05	42.54	92.58
nn M Net	50K	82.55	5.71	39.81	63.33
nn M Net †	50K	82.66	4.23	39.32	100.27
nn W Net	250K	84.01	7.05	42.54	92.58
nn M Net	250K	86.61	5.71	39.81	63.33
nn M Net †	250K	86.12	4.23	39.32	100.27

TABLE X: Ablation of Single versus Two-Stage Architectures

Panel of Figure 18 — Fig. 8: Qualitative results on SynMars-TW (top), SynMars-Air (middle), and MarsScapes (bottom).

Panel of Figure 19 — Fig. 9: Heatmap visualization of the encoder (top) and the bridges (down). The former focuses on local details, while the latter captures global context.

Panel of Figure 20 — Fig. 10: Heatmap visualization of the bridges in nn W Net (top) and our nn M Net (down). Our method shows greater contrast.

Semantic Segmentation

Semantic segmentation assigns a class label to every image pixel, enabling detailed scene understanding across domains and sensors.

See all 399 papers

To GAN or Not To GAN: Segmentation Analysis on Mars DEM
Jun 11, 2026Douglas Dziedzorm Agbeve, Aditya V. Handrale, Salim Fares +1Remote Sensing Image Segmentation Semantic Segmentation
University of Passau
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
Global-Local Contextual Progressive Expansion Network for Martian Landslide Segmentation in Multimodal Remote Sensing Imagery
Sep 11, 2026Leo Thomas Ramos, Sidike Paheding, Abel A. Reyes-Angulo +4Multi-Angular Remote Sensing Remote Sensing Image Segmentation
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 +3

Jun 11, 2026J/K move · Enter open · S save

To GAN or Not To GAN: Segmentation Analysis on Mars DEM

Douglas Dziedzorm Agbeve, Aditya V. Handrale, Salim Fares, Seif E. Idani

University of Passau

Remote Sensing Image Segmentation Semantic Segmentation Generative Adversarial Networks

Open paper AskPDF arXiv

To better understand Martian Surface, which is needed to enable Rovers navigate Mars with ease, it is necessary to be able to determine the location of mounds. Detecting and studying these morphologies can also help us find evidence of extraterrestrial life, in this case, more specifically, water or signs of life conducive environments. Detection of mounds was done by manually mapping morphological parameters onto Digital Elevation Models. This paper solves the problem by automatically detecting and or predicting mounds on Mars using Neural Network based Semantic Segmentation methodologies. This is done by using supervised semantic segmentation model and generative adversarial approach. A comparison of the approaches shows that adding extra artificially generated data did not improve the result.

Abstract

Semantic Segmentation

Explore similar work

To GAN or Not To GAN: Segmentation Analysis on Mars DEM

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

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