Glacial Landscape Change Evaluation Enabled by 3,340-Image Moraine Dataset and MCD-Net

Accurate segmentation of glacial moraines is crucial for understanding past glacial behaviour and assessing the impact of climate change on landscapes. Zhehuan Cao, Fiseha Berhanu Tesema, Ping Fu, Jianfeng Ren, and Ahmed Nasr from the University of Nottingham Ningbo China, have addressed the challenges of automated moraine mapping caused by poor image contrast and limited data availability. Their research presents the first large-scale dataset specifically for optical-only moraine segmentation, utilising over 3,300 manually annotated images of glaciated regions in China. The team developed MCD-Net, a computationally efficient deep learning model, which achieves impressive segmentation accuracy , a mean Intersection over Union of 62.3% and a Dice coefficient of 72.8% , while significantly reducing processing costs. This work establishes a new, reproducible benchmark and provides a readily deployable tool for monitoring glacial changes in high-altitude environments using only readily available imagery.

Optical Moraine Segmentation Dataset and Lightweight Network

This research introduces a large-scale optical-only moraine segmentation dataset, containing 3,340 manually annotated high-resolution images sourced from Google Earth, and focused on glaciated regions within Sichuan and Yunnan, China. To facilitate analysis, the researchers developed MCD-Net, a lightweight baseline model integrating a MobileNetV2 encoder, a Convolutional Block Attention Module (CBAM), and a DeepLabV3+ decoder. Comparative benchmarking against deeper backbone networks, specifically ResNet152 and Xception, reveals that MCD-Net achieves a mean Intersection over Union (mIoU) of 62.3% and a Dice coefficient of 72.8%, with a reduction in computational cost exceeding 60%. Despite the inherent challenges of ridge delineation, limited by sub-pixel width and spectral ambiguity, the results demonstrate the efficacy of utilising optical imagery for moraine segmentation. The developed MCD-Net provides a computationally efficient and accurate method for identifying these glacial landforms, contributing to glaciological studies and remote sensing applications in mountainous regions.

Moraine Segmentation Dataset and MCD-Net Development

Glacial segmentation is crucial for understanding past glacier behaviour and assessing landscape changes driven by climate. This study addresses the challenges of automated mapping caused by poor image contrast and limited high-resolution digital elevation models by introducing a novel, large-scale dataset specifically for moraine segmentation, compiled from 3,340 manually annotated high-resolution images sourced from Google Earth, focusing on glaciated regions within Sichuan and Yunnan, China. Scientists developed MCD-Net, a lightweight convolutional neural network integrating a MobileNetV2 encoder, a Convolutional Block Attention Module (CBAM), and a DeepLabV3+ decoder. This architecture was deliberately chosen for its efficiency and reproducibility, and benchmarking experiments compared MCD-Net against deeper backbones, including ResNet152 and Xception, revealing an mIoU of 62.3% and a Dice coefficient of 72.8%, while simultaneously reducing computational cost by over 60%.

The team rigorously evaluated performance across diverse geomorphological and environmental conditions to ensure robustness. The study pioneered an optical-only approach to moraine segmentation, circumventing the need for auxiliary topographic data. This innovation enhances reproducibility and scalability, particularly in high-altitude environments where data quality can be inconsistent. The technique reveals that reliable moraine-body segmentation can be achieved using imagery alone, despite limitations in delineating fine ridges due to sub-pixel width and spectral ambiguity. To promote open science, the dataset and associated code were made publicly available via GitHub at https://github.com/Lyra-alpha/MCD-Net, establishing a reproducible benchmark for moraine-specific segmentation and providing a deployable baseline for monitoring glacial activity in high-altitude regions. This work contributes the new dataset, the compact MCD-Net model, a systematic comparison of network architectures, and a robustness evaluation, ultimately advancing the field of glacial geomorphology.

Moraine Segmentation Dataset and Deep Learning Model

Glacial landforms are critical archives of Earth’s climatic history, with moraine ridges serving as valuable markers of former glacier extents. Scientists have introduced the first large-scale dataset for moraine segmentation, comprising 3,340 manually annotated high-resolution images sourced from Google Earth, covering glaciated regions of Sichuan and Yunnan, China. This work addresses limitations in current moraine mapping techniques, which traditionally rely on labour-intensive field surveys and are often constrained by the availability of high-resolution digital elevation models. The research team developed MCD-Net, a lightweight deep learning model integrating a MobileNetV2 encoder, a Convolutional Block Attention Module (CBAM), and a DeepLabV3+ decoder, specifically designed for efficient moraine segmentation.

Experiments revealed that MCD-Net achieves a mean Intersection over Union (mIoU) of 62.3% and a Dice coefficient of 72.8% in segmenting moraine features, demonstrating a reduction in computational cost exceeding 60% when compared to deeper network backbones such as ResNet152 and Xception. These measurements confirm the efficiency of the proposed lightweight architecture without significant performance degradation. Although precise ridge delineation is still limited by sub-pixel width and spectral ambiguity, results demonstrate that optical imagery alone can reliably segment moraine bodies. The study establishes a reproducible benchmark for moraine-specific segmentation, offering a deployable baseline for high-altitude glacial monitoring. The dataset and associated code are publicly available, facilitating further research and collaboration within the scientific community. The model’s performance is robust across diverse geomorphological and environmental conditions, highlighting its potential for broader application, and delivering a scalable and efficient solution for automated moraine mapping.

Moraine Segmentation via Lightweight Deep Learning

This study successfully establishes a new, reproducible benchmark for segmenting moraines using only optical imagery. Researchers curated a substantial dataset of 3,340 annotated images from glaciated regions of Sichuan and Yunnan, China, and developed MCD-Net, a computationally efficient deep learning model based on a DeepLabV3+ variant. Results demonstrate reliable moraine body delineation is achievable through imagery alone, offering a scalable method for glacial monitoring. The work highlights the benefits of lightweight neural network architectures, showing they outperform deeper models in both accuracy and efficiency when applied to data-scarce geomorphological applications.

Furthermore, the integration of attention mechanisms proved beneficial specifically for these compact backbones, suggesting a need for careful consideration of architectural design. The authors acknowledge that ridge delineation is constrained by the resolution of available satellite data, and that more detailed interpretation will require integration with higher-resolution or multi-source data. Future research could explore the use of unmanned aerial vehicle imagery, domain-adaptive learning to address spectral variability, and the incorporation of data from digital elevation models or synthetic aperture radar. This work ultimately advances automated moraine monitoring and contributes to improved understanding of landscape change and climate impact in alpine environments.