Optical-Only Moraine Segmentation Dataset

• Optical-Only Moraine Segmentation Dataset is a comprehensive collection of 3,340 high-resolution, manually annotated satellite images for precise moraine delineation in southwestern China.
• It addresses challenges like weak optical contrast and annotation variability through expert labeling, data augmentation, and merging of sub-pixel ridge classes.
• Baseline models such as MCD-Net achieve an mIoU of 62.3% and a Dice coefficient of 72.8%, establishing a robust benchmark for future geomorphological segmentation research.

The Optical-Only Moraine Segmentation Dataset is the first large-scale collection of high-resolution annotated satellite imagery designed specifically for pixel-level segmentation of moraine bodies using only optical data. Developed to address challenges in reconstructing past glacier dynamics and monitoring climate-driven landscape change, the dataset comprises 3,340 manually annotated image–mask pairs from glaciated regions of Sichuan and Yunnan, China. It serves as a reproducible benchmark for moraine segmentation and supports the development and assessment of deep learning models under conditions of weak optical contrast and limited auxiliary geospatial data (Cao et al., 5 Jan 2026).

1. Dataset Composition and Geographic Distribution

The dataset consists of 3,340 images, each paired with a corresponding segmentation mask. Acquired via Google Earth Pro between 2020 and 2025, these images have a spatial resolution of 1,024 × 1,024 pixels, translating to a ground sampling distance of approximately 0.5–2.0 m/pixel. Spatial coverage includes the principal glaciated ranges of southwestern China—especially in the provinces of Sichuan and Yunnan, demarcated by 26°–32° N and 98°–104° E. Key subregions encompass Gongga Shan, Que’er Shan, Yulong Shan, and Meili Shan, with image elevations spanning roughly 2,800 m to 5,200 m. This geographic breadth ensures comprehensive coverage of cirque, valley, and piedmont moraines under various illumination, shadow, and vegetation regimes (Cao et al., 5 Jan 2026).

2. Annotation Methodology and Quality Assurance

Annotation was conducted independently by three expert geomorphologists employing standard GIS and annotation tools, including QGIS and LabelMe, directly on Google Earth imagery. The initial labeling ontology comprised three classes: background (0), moraine body (1), and moraine ridge (2). Due to the extreme sparsity of the ridge class (≈0.2% of pixels) and significant inter-annotator variability (±2 pixels for ridge crests), ridge annotations were merged into the moraine body class, resulting in a binary segmentation problem (background vs. moraine body). Discrepancies in annotation were resolved via majority voting. For external validation, a subset of annotations was compared to earlier maps (Fu et al. 2012), with a cross-dataset IoU of 0.313, highlighting both the improvements gained via updated imagery and persistent uncertainty in sub-pixel ridge delineation (Cao et al., 5 Jan 2026).

3. Dataset Statistics, Partitioning, and Data Augmentation

The dataset is characterized by pronounced class imbalance, as approximately 90.0% of all pixels are background, 9.8% represent moraine bodies, and 0.2% correspond to ridges (subsequently merged). Moraine body sizes follow a long-tailed distribution, ranging from a few hundred to several hundreds of thousands of pixels (covering 0.2%–60% of an image). The partitioning is as follows:

Split	Number of Images	Percentage of Total
Training	2,630	≈ 78.8%
Validation	≈ 262	≈ 7.8% (10% of train for early stopping)
Test	293	≈ 8.8%

To foster model generalization, images were subject to data augmentation during pre-training: random scaling ([0.5, 2.0]), horizontal/vertical flips, rotations (±30°), and Gaussian blur (random σ) (Cao et al., 5 Jan 2026).

4. Data Format, Structure, and Accessibility

All images and segmentation masks are stored as 1,024 × 1,024 pixel lossless PNGs. Each mask uses a single-channel uint8 encoding: 0 for background, 1 for moraine body. The directory structure is as follows:

MCD_dataset/ ├─ images/ │ ├─ train/ (2,630 PNGs) │ ├─ val/ (≈ 262 PNGs) │ └─ test/ (293 PNGs) └─ masks/ ├─ train/ ├─ val/ └─ test/

The dataset and corresponding baseline code are openly available at https://github.com/Lyra-alpha/[MCD-Net](https://www.emergentmind.com/topics/mcd-net) (Cao et al., 5 Jan 2026).

5. Evaluation Metrics and Baseline Results

Performance benchmarking employs pixel-level metrics tailored for datasets with high class imbalance. Let TP, FP, and FN represent the true positives, false positives, and false negatives for the moraine-body class:

• Intersection over Union (IoU) for class $k$:

$\mathrm{IoU}_k = \frac{TP_k}{TP_k + FP_k + FN_k}$

• Mean IoU (mIoU) over $K=2$ classes:

$$\mathrm{mIoU} = \frac{1}{K}\sum_{k=1}^{K}\frac{TP_k}{TP_k + FP_k + FN_k}$$

• Dice coefficient (F1 for the positive/moraine class):

$\mathrm{Dice} = \frac{2\,TP}{2\,TP + FP + FN}$

On the held-out test set, the MCD-Net baseline utilizing a MobileNetV2 encoder, Convolutional Block Attention Module (CBAM), and DeepLabV3+ decoder, achieves an mIoU of 62.3% and a Dice coefficient of 72.8%. Comparative baselines using ResNet152 and Xception without CBAM achieve lower mIoU and Dice scores (60.2% / 70.8% and 56.5% / 66.2%, respectively), despite having significantly greater model parameters (74.99M/54.7M) and computational cost (433.9/333.7 GFLOPs), compared to MCD-Net's ≈5.83M parameters and 105.7 GFLOPs (Cao et al., 5 Jan 2026).

6. Challenges, Limitations, and Prospects

Sub-pixel ridge delineation remains problematic: many ridges are narrower than 3 pixels, leading to notable inter-annotator variability and low IoU for isolated crest features. Spectral ambiguity introduced by vegetation, shadows, or artifacts can cause confusion between moraine and background, especially for degraded or very small deposits (<500 pixels). Future enhancements may include multi-modal fusion with lightweight DEM or SAR inputs, advanced domain adaptation to correct inter-scene illumination or color differences, UAV-based photogrammetry for sub-meter ridge tracing, and more sophisticated small-object/attention mechanisms to improve detection of isolated moraines. The dataset’s design and open accessibility facilitate subsequent methodological improvements and reproducible benchmarking for high-altitude geomorphological segmentation tasks (Cao et al., 5 Jan 2026).