MineC2FNet: Coarse-to-Fine Segmentation

• The paper introduces MineC2FNet, a coarse-to-fine domain incremental learning framework that leverages abundant coarse labels to improve fine-grained mining footprint segmentation in multispectral imagery.
• It employs a teacher-student architecture with attentive distillation at both feature and prediction levels to effectively mitigate domain shift and boundary imprecision.
• Experimental results demonstrate superior performance with a mIoU of 73.64%, and ablation studies confirm that feature fusion components like gated fusion and CBAM significantly boost segmentation accuracy.

MineC2FNet is a coarse-to-fine domain incremental learning framework for mining footprint segmentation in multispectral satellite imagery when coarse boundary labels are abundant and precise, fine-grained annotations are scarce. It is designed to exploit abundant coarse data to enhance fine-grained mining footprint segmentation under severe domain shift, using a teacher-student architecture with attentive distillation at both the feature and prediction levels. The framework was introduced in "Coarse-to-Fine Domain Incremental Learning with Attentive Distillation for Mining Footprint Segmentation in Multispectral Imagery" (Handoyo et al., 23 May 2026).

1. Problem setting and domain shift

Mining footprint segmentation requires delineating pits, waste rock, tailings, ponds, and infrastructure. Fine-grained, pixel-accurate labeling requires domain expertise and is expensive to produce, so most available global data use coarse polygons with smoothed outlines and merged substructures. Training on such coarse labels propagates boundary imprecision into model predictions. MineC2FNet addresses this setting by treating coarse and fine annotation regimes as distinct but related domains rather than assuming that direct transfer is benign (Handoyo et al., 23 May 2026).

The coarse and fine domains differ in class distributions and geometry. The work quantifies boundary roughness via

$R = \frac{P^2}{4\pi A},$

where $P$ is perimeter and $A$ is area. Coarse masks peak at $R \approx 2$ while fine masks peak up to $R \approx 12$. Coarse annotations also present a more balanced mining/non-mining proportion, while fine labels exhibit severe mining sparsity, with non-mining $>80\%$. The reported shift arises from label imprecision, class imbalance, geographic and sensor variation. This combination makes naïve transfer from coarse to fine detrimental, particularly around boundaries.

A plausible implication is that MineC2FNet is not simply a refinement network appended to a standard segmentation backbone. Its formulation depends on the explicit recognition that generalized mining cues and boundary-specific cues should be transferred differently.

2. Formal formulation and architectural design

MineC2FNet formulates coarse-to-fine segmentation as a two-task domain incremental learning problem with attentive distillation. Let

$D_c = \{(X_i^c, Y_i^c)\}_{i=1}^{N_c}$

be the coarse dataset and

$D_f = \{(X_i^f, Y_i^f)\}_{i=1}^{N_f}$

the fine dataset. The model $M$ has teacher $M(\theta_T)$ and student $P$0. The optimization for Task 2 is

$P$1

Task 1 trains a teacher on coarsely labeled data to capture generalized patterns of mining footprints. Task 2 freezes the teacher and trains a student on the fine dataset. The student refines boundaries using fine labels while selectively distilling knowledge from the teacher (Handoyo et al., 23 May 2026).

The backbone/decoder is a Feature Pyramid Network with DenseNet-121 backbone, initialized on ImageNet. Inputs are multispectral with 6 channels (RGB, NIR, SWIR) at $P$2 resolution. Teacher and student are both FPN-DenseNet-121, and the teacher is frozen in Task 2. During Task 2, an identical frozen copy of the teacher backbone runs in parallel with the trainable student backbone on the same fine input $P$3, producing coarse-domain generalized features $P$4 and fine-domain specific features $P$5.

Attentive Feature Injection is implemented as a sequence of feature fusion and refinement operations. First, Gated Fusion uses a learnable, pixel-wise gating mechanism to merge the two streams, enabling adaptive mixing of general context from the teacher and precise details from the student. Second, CBAM applies sequential channel and spatial attention on the fused features, enhancing salient regions and feature channels. Third, Diffusion Refinement iteratively sharpens features through a stack of refinement blocks:

$P$6

$P$7

$P$8

The final refined feature map $P$9 feeds the segmentation head.

This design decomposes knowledge into generalized, domain-agnostic cues and domain-specific boundary details. The teacher captures the former from abundant coarse data, while the student learns the latter from scarce fine labels.

3. Loss functions and attentive distillation

The Task 2 objective is a sum of the student’s segmentation loss and the distillation loss:

$A$0

The student loss is a multi-term segmentation objective,

$A$1

combining region overlap, structural similarity, and explicit boundary emphasis. Boundary-aware weighting uses a Sobel-based importance map. Let $A$2 be Sobel gradients of $A$3 and

$A$4

After min-max normalization,

$A$5

The weighted boundary loss is

$A$6

where $A$7 is the number of pixels (Handoyo et al., 23 May 2026).

Prediction-level Attentive Knowledge Transfer uses KL-divergence distillation with temperature $A$8, but only where the teacher is better than the student, determined by BCE comparisons against fine labels. Let $A$9 and $R \approx 2$0 denote the teacher and student outputs for image $R \approx 2$1, and $R \approx 2$2 and $R \approx 2$3 the corresponding softened distributions. The method defines image-level and pixel-level selection masks and combines them into a hybrid selection mask:

$R \approx 2$4

The final attentive distillation term is

$R \approx 2$5

with $R \approx 2$6 for numerical stability.

Feature-level selectivity is realized through architecture: the frozen teacher backbone provides generalized features; Gated Fusion learns pixel-wise gates to combine teacher and student features; CBAM computes channel-wise and spatial attention maps; Diffusion Refinement progressively sharpens features. Prediction-level selectivity is achieved by confidence-based masking, so distillation is applied only for images and pixels where the teacher is more accurate than the student. This prevents negative transfer from the teacher’s imprecise boundaries.

Multispectral bands (RGB, NIR, SWIR) are concatenated as input to both teacher and student backbones. Attention is not explicitly band-aware, but ablations show that adding NIR and SWIR is critical for discriminating mining materials, including altered rocks, tailings, and water ponds.

4. Datasets, curation, and training protocol

MineC2FNet was developed and evaluated with a coarse dataset derived from Global-scale mining polygons (Version 2) by Maus et al. and a new, expertly validated fine dataset. The coarse dataset begins from 44,929 polygons globally. Imagery comes from Sentinel-2 (10 m) and Landsat 8 (15/30 m) with RGB, NIR, and SWIR bands. After filtering polygons by minimum area threshold $R \approx 2$7—the smallest mine in the fine dataset—1,380 coarse-labeled images were retained. Labels are binary (Mining vs Non-Mining) with rough edges and merged features (Handoyo et al., 23 May 2026).

The fine dataset contains 219 images with precise mining boundaries from globally distributed sites, harmonized from multiple sources following published protocols. Bands are RGB and NIR; 200 images also include SWIR. Resolution and extent vary across Sentinel-2 and Landsat 8 imagery. The split is $R \approx 2$8 into train/val/test, corresponding to 153/22/44 images. The set is balanced across commodities, regions, and five Köppen-Geiger climate zones, while still exhibiting severe class imbalance with $R \approx 2$9 non-mining.

Dataset	Size and source	Key properties
Coarse dataset	1,380 retained images from Global-scale mining polygons (Version 2) by Maus et al.	Binary labels, rough edges, merged features, Sentinel-2 and Landsat 8, RGB/NIR/SWIR
Fine dataset	219 expertly validated images	Precise boundaries, global distribution, 153/22/44 split, severe class imbalance, five Köppen-Geiger climate zones

The training pipeline follows two stages. Task 1 uses standard supervised training with Binary Cross-Entropy on $R \approx 12$0 to learn generalizable features under coarse supervision. Task 2 freezes teacher parameters and trains the student on $R \approx 12$1 with the composite loss described above. Optimization uses AdamW with learning rate $R \approx 12$2, weight decay $R \approx 12$3, batch size 8, for up to 100 epochs. A 5-epoch linear warm-up precedes a ReduceLROnPlateau schedule with factor 0.5 on plateau of validation mIoU for 3 epochs, and EarlyStopping on validation mIoU with patience 10 saves the best checkpoint. Data augmentation includes random flips, 90-degree rotations, shifts, scaling (limit 0.3), rotation ($R \approx 12$4), and shift limit 0.1. Implementation uses Python 3.9, TensorFlow 2.13, and an NVIDIA RTX A4000 GPU. The dataset and code are publicly available at the reported repository.

5. Evaluation results and empirical behavior

Evaluation uses Pixel Accuracy, Mean F1 (mF1), and Mean IoU (mIoU). Baselines include transfer learning with U-Net, DeepLabV3+, FPN, and Prithvi; domain adaptation methods UDAforRS and BUS; continual learning methods LwF, LwM, and Replay; class-incremental methods SPPA and CCDA; and domain-incremental methods MDIL-SS and GSMF-RS-DIL. Extensive experiments against these state-of-the-art approaches demonstrate that MineC2FNet achieves superior performance while effectively handling domain shift (Handoyo et al., 23 May 2026).

On the fine test set, MineC2FNet achieves Accuracy 92.33%, mF1 84.10%, and mIoU 73.64%, with the best mIoU among all baselines. Selected comparison points are listed below.

Method	mIoU on fine test set	Note
MineC2FNet	73.64%	Best mIoU among all baselines
MDIL-SS	71.99%	Domain-incremental baseline
GSMF-RS-DIL	71.25%	Domain-incremental baseline
FPN	68.26%	Transfer learning baseline
DeepLabV3+	67.47%	Transfer learning baseline
U-Net	67.26%	Transfer learning baseline

The framework improves mIoU by +5.4 points over the FPN baseline (0.7364 vs 0.6825) with a modest inference overhead (11.00 FPS vs 13.38 FPS). Climate robustness analysis shows mIoU up to 77.34% in temperate climates and competitive performance in arid and tropical zones; performance dips in underrepresented cold/polar zones, with polar mIoU 69.42%. Qualitative results indicate finer, more realistic boundaries than the baselines. When applied to the coarse dataset, MineC2FNet outputs finer contours than the coarse labels, indicating effective refinement even outside the fine domain.

A plausible implication is that the method benefits not only from transferring generalized mining signatures but also from actively suppressing boundary noise introduced by coarse supervision.

6. Ablation findings, limitations, and related naming

Ablation analyses isolate the contribution of each component. For Attentive Feature Injection, Gated Fusion alone provides gains, and adding CBAM and Diffusion Refinement yields substantial improvements, with the best mIoU of 68.41% when all components are combined in the RGB+NIR experiment. For Attentive Knowledge Transfer, standard unconditional distillation yields only marginal gains over no distillation (mIoU 62.59 vs 62.35), whereas image-level selection improves to 67.17, pixel-level selection to 67.84, and hybrid selection to 67.91. These results confirm the importance of selective distillation to avoid negative transfer from noisy coarse boundaries (Handoyo et al., 23 May 2026).

Band ablations show that RGB+NIR increases mIoU from 57.09 to 62.35, and incorporating SWIR yields 67.29. Coarse data curation also matters: pretraining on a curated coarse subset aligned to the fine data’s minimum area outperforms using the full noisy coarse set. Confusion analysis reports reduced false positives and increased true positives for the mining class relative to strong domain-incremental baselines. The main sensitivity lies in the selection masks defined by BCE comparisons and in the strength of boundary loss, while the computational overhead of GF+CBAM+DR is described as modest for the observed gains.

The reported limitations are explicit. The approach currently targets binary Mining vs Non-Mining segmentation, and extension to multi-class mining typologies is future work. Effectiveness depends on teacher quality: if the teacher overfits to coarse biases, selective distillation reduces but does not eliminate negative transfer. Remaining domain shift challenges persist in underrepresented regions and sensors, and generalization to other remote sensing tasks is described as promising but requiring task-specific boundary losses and attention tuning. Ethical considerations are also stated: improved mapping can support socio-environmental monitoring and accountability but must be used responsibly, respecting local communities and data governance.

The name MineC2FNet can be confused with "C^2F-Net," but the two are distinct. The paper "Camouflaged Object Detection via Context-aware Cross-level Fusion" does not introduce a model named MineC2FNet; it presents C^2F-Net for camouflaged object detection, with modules such as ACFM, DGCM, and CIM, and an expanded journal version that adds the Camouflage Inference Module and low-level refinement guided by a coarse prediction (Chen et al., 2022). This suggests that the shared "coarse-to-fine" naming motif should not be interpreted as architectural identity or direct lineage without additional evidence.