LEMMA: Laplacian pyramids for Efficient Marine SeMAntic Segmentation

Ishaan Gakhar; Laven Srivastava; Sankarshanaa Sagaram; Aditya Kasliwal; Ujjwal Verma

cs.CVMar 26, 2026

LEMMA: Laplacian pyramids for Efficient Marine SeMAntic Segmentation

Authors: Ishaan Gakhar, Laven Srivastava, Sankarshanaa Sagaram, Aditya Kasliwal, Ujjwal Verma

Organizations: Manipal Institute of Technology, Manipal Academy of Higher Education, India

Abstract

Semantic segmentation in marine environments is crucial for the autonomous navigation of unmanned surface vessels (USVs) and coastal Earth Observation events such as oil spills. However, existing methods, often relying on deep CNNs and transformer-based architectures, face challenges in deployment due to their high computational costs and resource-intensive nature. These limitations hinder the practicality of real-time, low-cost applications in real-world marine settings. To address this, we propose LEMMA, a lightweight semantic segmentation model designed specifically for accurate remote sensing segmentation under resource constraints. The proposed architecture leverages Laplacian Pyramids to enhance edge recognition, a critical component for effective feature extraction in complex marine environments for disaster response, environmental surveillance, and coastal monitoring. By integrating edge information early in the feature extraction process, LEMMA eliminates the need for computationally expensive feature map computations in deeper network layers, drastically reducing model size, complexity and inference time. LEMMA demonstrates state-of-the-art performance across datasets captured from diverse platforms while reducing trainable parameters and computational requirements by up to 71x, GFLOPs by up to 88.5%, and inference time by up to 84.65%, as compared to existing models. Experimental results highlight its effectiveness and real-world applicability, including 93.42% IoU on the Oil Spill dataset and 98.97% mIoU on Mastr1325.

Figures & tables

Figure 1: Visualization of the Laplacian pyramid of depth 2 of an RGB image. Vital spatial and structural information is present at each level, notably in the edges available at different resolutions. Here, $L_{1}$ , $L_{2}$ and Residual ( $L_{3}$ ) refer to the first, second and last layer of the decomposed Laplacian Pyramid.

Figure 2: A schematic overview of the proposed model-LEMMA. The sections highlighted in green, blue, and yellow symbolize the LFB, MFB, and HFB, respectively. $L_{1}$ , $L_{2}$ , and Residual ( $L_{3}$ ) represent the three layers of the decomposed Laplacian pyramid. The ’nc’ for each dataset is the total number of classes in the dataset. Each residual block chain has the corresponding number of blocks, as explained in Section 3 .

Model	mIoU	#Params (M)	GFLOPs	Inference Time (ms)
DeepLabv3 + Xception [ 10 ]	80.65	54.71	124.65	29.83
ERFNet [ 40 ]	81.68	2.06	22.16	15.07
Pretrained SegNet [ 2 ]	81.80	29.45	241.29	22.31
CFPNet [ 32 ]	82.28	0.54	5.85	35.61
ERFNetv2 [ 34 ]	82.45	1.23	4.18	18.33
DABNet [ 26 ]	82.64	0.75	7.76	15.59

Table 1: Comparison of various models with LEMMA on mIoU (in percent) across MaSTr1325 [ 4 ] and Oil Spill Drone [ 42 ] datasets. Parameters, GFLOPs, and inference time are provided for both ResNet-50 and ResNet-101 backbones where applicable. Results for LEMMA are of residual block (Section 3 ) configurations 7,7,1 and 6,7,4 respectively (separated by forward slashes for Mastr1325 and Oil Spill). Values for SOTA comparisons separated by slashes refer to R50/R101 variants. Unreported values in original works or lack of publically available weights have been left blank.

ResNet-50 (R) Backbones
Model	mIoU	#Params (M)
R-LinkNet [ 7 ]	90.06	-
R-SGDBNet [ 17 ]	90.22	-
R-MANet [ 18 ]	91.92	26
R-FPN [ 28 ]	91.95	28
R-DeepLabv3 [ 9 ]	92	39.6

Table 2: Comparison of various models with LEMMA on mIoU across the Oil Spill Drone [ 42 ] dataset. Parameters are provided for both ResNet-50 (R) and Efficient-Net-b1 backbones (E) where applicable. Results for LEMMA are of residual block (Section 3 ) configurations 6,7,4. Missing data is due to unreported values in original works or lack of publicly available weights. Parameters for some of the models have been calculated by the parameter count of backbones and other modules of the architecture as made available by authors.

Figure 3: Visualization of the qualitative outputs on the Mastr1325 Dataset.

Figure 4: Visualization of the qualitative outputs on the Oil Spill Drone Dataset.

Dataset	NRB L	NRB M	NRB H	mIOU
Mastr1325	3	3	3	0.9864
	3	4	3	0.9871
	3	5	3	0.9876
	3	6	3	0.9876
	3	7	3	0.9886
	4	7	3	0.9882

Table 3: Impact of Number of residual blocks in each branch of the model and the corresponding results on both datasets. The best results for every branch are bolded.

Dataset	Loss Function	mIOU
MaSTr1325	Focal	0.9897
	Dice	0.9872
	CE+Dice	0.9886
Oil Spill	Focal	0.9342
	Dice	0.9262
	CE+Dice	0.9294

Table 4: Impact of various loss functions with best configurations of residual blocks on both datasets. The best loss function for each dataset is bolded.

Semantic Segmentation

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

See all 414 papers

MambaMPD: A Mamba-Driven Segmentation Framework for Marine Pollution Detection from Remote Sensing Imagery
Sep 14, 2026Shuaiyu Chen, Wei Han, Peng Ren +2Remote Sensing Image Segmentation Visual SSMs
Department of Computer Science, University of Exeter, Exeter, United Kingdom · School of Computer Science, China University of Geosciences, Wuhan, China · College of Oceanography and Space Informatics, China University of Petroleum (East China), Qingdao, China
GeoSAM-Lite: A Lightweight Foundation Model for Onboard Remote Sensing Segmentation
Jul 4, 2026Yongcong Wang, Jie Zhang, Rui Jiang +3Remote Sensing Image Segmentation Geospatial Foundation Models
College of Telecommunications and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, China
OceanMAE: A Foundation Model for Ocean Remote Sensing
Apr 9, 2026Viola-Joanna Stamer, Panagiotis Agrafiotis, Behnood Rasti +1Remote Sensing Self-Supervised Visual Representation Learning
Faculty of Electrical Engineering and Computer Science, Technische Universität Berlin, Germany and with the Berlin Institute for the Foundations of Learning and Data (BIFOLD), 10623 Berlin, Germany.

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

MambaMPD: A Mamba-Driven Segmentation Framework for Marine Pollution Detection from Remote Sensing Imagery

Shuaiyu Chen, Wei Han, Peng Ren, Chunbo Luo+1

Department of Computer Science, University of Exeter, Exeter, United Kingdom · School of Computer Science, China University of Geosciences, Wuhan, China · College of Oceanography and Space Informatics, China University of Petroleum (East China), Qingdao, China

Remote Sensing Image Segmentation Visual SSMs Boundary-Aware Segmentation

Open paper AskPDF arXiv

Accurate marine pollution detection (MPD) is essential for protecting coastal ecosystems and marine biodiversity. Vision Mamba models have shown promise in remote-sensing semantic segmentation by efficiently capturing long-range dependencies and global context, yet their potential for MPD remains underexplored. MPD is particularly challenging because of low signal-to-noise ratios, fragmented pollution patterns, and indistinct boundaries caused by the visual similarity between pollutants and the surrounding sea. To address these issues, we propose MambaMPD, an enhanced Mamba-based framework incorporating two complementary structural priors: Frequency-Aware Augmentation (FAA) and multi-scale Edge-Guided Attention (EGA). FAA integrates wavelet transforms into the encoder to decompose features into multi-scale frequency subbands, enabling the model to capture low-frequency contextual semantics and high-frequency structural details needed to identify small, low-contrast, and irregular pollution patterns. EGA adaptively fuses hierarchical, Laplacian-derived boundary cues with deep semantic representations, refining encoder features before decoding to sharpen boundaries and reduce ambiguity in visually confusing, spatially fragmented scenes. Together, these modules improve sensitivity to subtle pollution signals while preserving fine boundary structures. A U-Net-style decoder with squeeze-and-excitation attention and deep supervision progressively restores and refines semantic and spatial information across scales. Extensive experiments on two benchmark MPD datasets show that MambaMPD achieves higher mIoU than competing methods while requiring substantially less computation than foundation-model-based approaches. On MADOS, it improves F1 by 3.6% over OSDMamba; on M4D, it raises Oil Spill IoU by 6.82% over TransOilSeg.

Abstract

Semantic Segmentation

Explore similar work

MambaMPD: A Mamba-Driven Segmentation Framework for Marine Pollution Detection from Remote Sensing Imagery

GeoSAM-Lite: A Lightweight Foundation Model for Onboard Remote Sensing Segmentation

OceanMAE: A Foundation Model for Ocean Remote Sensing