Coarse-to-Fine Domain Incremental Learning with Attentive Distillation for Mining Footprint Segmentation in Multispectral Imagery

arXiv:2605.24460 · cs.CV, cs.AI · Submitted 2026-05-23 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Coarse-to-Fine Domain Incremental Learning with Attentive Distillation for Mining Footprint Segmentation in Multispectral Imagery".

Jane: The gist The proposed MineC2FNet is a coarse-to-fine domain incremental learning framework that exploits abundant coarse data to enhance fine-grained mining footprint segmentation by adopting an attentive distillation architecture.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper titled Coarse-to-Fine Domain Incremental Learning with Attentive Distillation for Mining Footprint Segmentation in Multispectral Imagery. Jane, what does that title tell us about what they are actually trying to tackle?

Jane: It tells us they are dealing with a problem where you have tons of data that is coarse, meaning the boundaries are broad, and you want to use that information to get better at segmenting things when the data gets much finer. It’s about incremental learning across different domains, which is tricky because the data shifts between those domains.

Lu: Exactly. The key idea here is moving from broad labels in one domain to very precise boundaries in another, and they are proposing a framework that uses what they call attentive distillation to bridge that gap between the coarse and fine data.

Meng: So, instead of just training on the fine data alone, they want to leverage the abundant coarse data while still getting those high-quality, expert-validated boundaries from the fine set.

Lalam: What I see is a two-task setup: Task one is on a coarse domain like general pavement or roads, and Task two is on a fine domain with precise labels like sidewalks or vegetation. The goal is to train one model that uses the knowledge from the first task to perform better on the second task.

Tom: Right, so it’s not just about getting a good segmentation in one domain; it’s about making sure that general knowledge from the coarse data transfers correctly and selectively to help with those fine details.

The paper's summary: Jane: Okay, so if we look at the summary of Coarse-to-Fine Domain Incremental Learning with Attentive Distillation for Mining Footprint Segmentation in Multispectral Imagery, they are proposing a specific framework called MineC2FNet.

Tom: MineC2FNet is described as a coarse-to-fine domain incremental learning framework that exploits abundant coarse data to enhance fine-grained mining footprint segmentation by adopting an attentive distillation architecture. That’s the core mechanism they are putting forward, right?

Lu: They frame it as a domain incremental learning problem with two sequential tasks, Task one on the Coarse Domain and Task two on the Fine Domain. The goal is to train a model that leverages knowledge from the coarse domain to perform better in the fine domain, but they specifically focus on selective attentive distillation to transfer only relevant general knowledge while filtering out noise from the coarse task.

Meng: They use a composite loss function for Task two, which combines a student loss and a distillation loss. The student loss has four components: Binary Cross-Entropy, Dice Loss, Structural Similarity Loss, and Boundary Loss.

Lalam: And the distillation part is formulated as Kullback-Leibler divergence between the softened probability distributions of the teacher and student models, scaled by a temperature factor T. That’s how they try to transfer knowledge from teacher to student.

Tom: So, it’s not just one loss function; they are combining traditional segmentation losses with this distillation term to guide the learning process on the fine domain. What does that mean for the practical training?

The paper's improvements: Jane: They suggest a few key improvements to this approach, focusing on how they make that knowledge transfer work better. One big thing is introducing an attentive feature injection pipeline for domain bridging.

Lu: They’ve got a teacher feature extractor, which is just a frozen copy of the teacher’s backbone, and a student feature extractor with trainable parameters. This lets the student model have two sets of features: generalized source domain features from the coarse data and specific target domain features from the fine data.

Tom: That sounds like they’re trying to keep both types of information available simultaneously during training, which is smart for handling that domain shift you mentioned earlier.

Meng: They also use gated fusion, which uses a technique from Takikawa et al., two thousand nineteen to dynamically merge contextual information from the coarse teacher backbone with the high-frequency details from the fine student backbone <ref:2605.24460#pg2>. It allows the network to decide which feature source is more important at any given point.

Lalam: Then they have Convolutional Block Attention Module, or CBAM, which refines those fused features by looking at both channel and spatial attention maps to find what features are most salient in terms of channels and where the model should focus its attention spatially.

Jane: And finally, they add Diffusion Refinement. This is a new module that progressively enhances the features by running them through a sequence of custom refinement blocks, combining a convolution, a CBAM layer, and a residual connection to sharpen those noisy features.

Tom: That sounds like they are building this sophisticated feature transfer step-by-step: injection, merging with gating, attention refinement via CBAM, and then final sharpening with diffusion. It’s quite complex architecture for incremental learning.

Conclusion: Jane: So, to wrap up the Coarse-to-Fine Domain Incremental Learning with Attentive Distillation for Mining Footprint Segmentation in Multispectral Imagery paper, the main point is that it successfully uses coarse data to improve fine segmentation by selectively transferring knowledge.

Tom: They demonstrate effectiveness by achieving a mIoU of seventy-three point six four percent and an accuracy of ninety-two point three three percent compared to other methods across five learning categories. And they specifically highlight that including the Boundary Loss yielded the highest performance at zero point six six eight seven mIoU because it weighted optimization toward edge pixels, which makes sense for segmentation tasks where edges matter most.

Lu: The paper also shows a reduction in False Positive rate to three point five two percent, which is about twenty-nine to forty percent less than what they found in MDIL-SS or GSMF-RS-DIL, indicating their selection mechanism for distillation is working effectively.

Meng: The optimal input size they found was two hundred fifty-six by two hundred fifty-six pixels, and it achieved a slightly higher testing F1-score of zero point seven five nine six compared to the previous model's score of zero point seven five three nine, which means they optimized the input resolution for their specific fine-tuned model.

Lalam: What this all means for us is that we have a novel framework that can effectively handle the domain shift in remote sensing data by intelligently filtering what information to transfer between domains, and they are also introducing a new globally distributed dataset of two hundred nineteen images to help researchers move forward <ref:2605.24460#pg2>.

Tom: So, Coarse-to-Fine Domain Incremental Learning with Attentive Distillation for Mining Footprint Segmentation in Multispectral Imagery is a framework that proves you can use massive amounts of coarse data to get state-of-the-art results when you have fine data available. What's next?

Jane: Next, we’ll talk about how this works on diverse climate zones, which shows the model has global robustness, achieving a peak mIoU of seventy-seven point three four percent in temperate climates.

Lu: That robustness across five Koppen–Geiger climate zones is interesting because it proves the model can effectively differentiate mining footprints across different geographic locations.

Meng: I just wonder about deployment, though; how complex is this architecture to run when you try to deploy it at scale?

Lalam: MineC2FNet is quite robust in its design, and they are planning to extend this framework to complex multiclass segmentation of diverse mining and non-mining categories.

Monash University Indonesia · Monash University Australia · Northeastern University China · The University of Queensland Australia

cs.CV, cs.AI

Submitted: 2026-05-23

Updated: 2026-05-28

Comments: Accepted at the 35th International Joint Conference on Artificial Intelligence (IJCAI 2026), AI and Social Good track

DOI: 10.24963/ijcai.2026/800

Code: https://github.com/risqiutama/MineC2FNet

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: The gist The proposed MineC2FNet is a coarse-to-fine domain incremental learning framework that exploits abundant coarse data to enhance fine-grained mining footprint segmentation by adopting an

Key concepts

Coarse-to-Fine Domain Incremental Learning: This is a learning approach where a model first learns on large, general (coarse) data and then adapts that knowledge to a smaller, more detailed (fine) dataset. The goal is to leverage the broad information from the coarse data to help accurately segment specific details in the fine data without being confused by domain-specific noise.
Attentive Feature Injection: This technique involves injecting features from two feature extractors into a student model. One extractor is frozen (teacher), providing generalized knowledge, and the other is trainable (student), capturing specific target domain details. A Gated Fusion module then intelligently merges these two feature sources based on their contextual relevance.
Hybrid Distillation Loss: Instead of standard knowledge transfer, this loss uses a selection mechanism to decide when to distill information. It checks if the teacher's prediction is better than the student's at both the image and pixel levels before applying distillation, ensuring only relevant high-quality knowledge is transferred.
Boundary Loss (LBoundary): This specific loss component focuses optimization on correctly classifying pixels located at the edges or boundaries of segmented objects. By weighting these edge pixels more heavily during training, the model is forced to produce sharper and more accurate segmentation lines.

Terminology

Distillation — 2× in this paper · explained in 2 episodes
Domain Shift — 2× in this paper · explained in 7 episodes
Adaptation — 1× in this paper · explained in 1 episode
Domain Adaptation — 1× in this paper · explained in 5 episodes

Summary

The gist The proposed MineC2FNet is a coarse-to-fine domain incremental learning framework that exploits abundant coarse data to enhance fine-grained mining footprint segmentation by adopting an attentive distillation architecture.

Problem Formulation

The problem is framed as a domain incremental learning problem consisting of two sequential tasks: Task 1 on the Coarse Domain Dc and Task 2 on the Fine Domain Df, where the goal is to train a model M(θ) that leverages knowledge from Dc to excel at Df, requiring selective attentive distillation to transfer relevant general knowledge while discarding domain-specific noise from the coarse task

Coarse-to-Fine Domain Incremental Learning

The training for Task 1 follows standard semantic segmentation procedures using Binary CrossEntropy (BCE) as the loss function For Task 2, the training process is orchestrated by a composite loss function, Ltotal, which combines Lstudent and Ldistill

Student Loss: The student loss, Lstudent, is a comprehensive objective function designed to train the student model M(θS) on the fine-grained dataset Df using four components: Binary Cross-Entropy (LBCE), Dice Loss (LDice), Structural Similarity Loss (LSSIM), and Boundary Loss (LBoundary)
Distillation Loss: The distillation loss, Ldistill, is formulated as the standard Kullback-Leibler (KL) divergence between the softened probability distributions of the teacher (P T) and student (Q T), scaled by a temperature factor T

Attentive Feature Injection

The framework employs an attentive feature injection pipeline to intelligently transfer generalized feature extraction capabilities from the teacher model M(θT) to the student M(θS) This multi-stage pipeline consists of four stages: Feature Injection, Gated Fusion (GF), Convolutional Block Attention Module (CBAM), and Diffusion Refinement (DR)

** Feature Injection involves injecting a subset of M(θT), particularly the feature extractor, to M(θS), enabling dual, identical feature extractors: a ”teacher feature extractor” (a frozen copy of the teacher’s backbone) and a ”student feature extractor” (with trainable parameters) **

** Gated Fusion uses [Takikawa et al., 2019] to dynamically merge contextual information from the coarse teacher backbone with high-frequency details from the fine student backbone, allowing the network to adaptively decide the contribution of each feature source **

** CBAM refines fused features by sequentially inferring attention maps along both channel and spatial dimensions to learn what features are most salient (channel-wise) and where to focus its attention (spatial-wise) **

** Diffusion Refinement introduces a new module that progressively enhances features by propagating information through a sequence of custom refinement blocks, combining a convolution, a CBAM layer, and a residual connection to effectively sharpen noisy features **

Attentive Knowledge Transfer

An attentive knowledge transfer mechanism is introduced at the prediction level to selectively apply distillation only when the teacher’s prediction is more accurate than the student’s prediction This involves two layers of selection:

Image-level Selection: An indicator function I(img) returns 1 if LBCE(Y f i, Pi) < LBCE(Y f i, Qi) and 0 otherwise, and the distillation loss is formulated as Limage distill
Pixel-level Selection: An indicator function I(pxl) returns 1 if LBCE(Y f ij, Pij) < LBCE(Y f ij, Qij) and 0 otherwise, and the distillation loss is formulated as Lpixel distill

The final distillation loss is computed using a hybrid selection mask I hybrid by performing element-wise multiplication of I(pxl) and I(img)

Results and Discussion

MineC2FNet demonstrates superior performance, achieving the highest Accuracy (92.33%) and mIoU (73.64%) compared to different methods across five distinct learning categories The framework successfully produces fine details while preserving coarse general knowledge, as shown in Figure 5(d) The final inclusion of Boundary Loss yielded the highest performance (0.6687 mIoU) by weighting optimization toward edge pixels MineC2FNet restricts the False Positive (FP) rate to 3.52%, marking a reduction of approximately 29–40% relative to MDIL-SS (4.96%) and GSMF-RS-DIL (5.89%) The optimal input size for the fine-tuned model was determined to be 256 × 256, achieving a slightly higher testing F1-score (0.7596 vs 0.7539)

Conclusion

MineC2FNet is a novel coarse-to-fine domain incremental framework that proves highly effective, achieving a state-of-the-art mIoU of 73.64% by successfully selectively transferring knowledge from coarse to fine-grained data This work also introduced a new, expertly validated global dataset to facilitate further research in this domain Future work will extend this framework to complex multiclass segmentation of diverse mining and non-mining categories

Acknowledgments

This study was supported in part by the Google Research Scholar Program, Australian Research Council LP200301160, and the Ford Foundation The work of A. T. H. was partially supported by the Monash University Indonesia PhD Tuition Scholarship We also sincerely thank John R. Owen for his contribution to the development of the training data

References

[Alfarra et al., 2024] Motasem Alfarra, Zhipeng Cai, Adel Bibi, Bernard Ghanem, and Matthias Muller. Simcs: Simulation for domain incremental online continual segmentation. Proceedings of the AAAI Conference on Artificial Intelligence, 38(10):10795–10803, Mar. 2024

[Ang et al., 2023] Michelle Li Ern Ang, John R. Owen, Christopher N. Gibbins, El ´ eonore L ´ ebre, Deanna Kemp, Muhamad Risqi U Saputra, Jo-Anne Everingham, and Alex M Lechner. Systematic review of gis and remote sensing applications for assessing the socioeconomic impacts of mining. The Journal of Environment & Development, 32(3):243–273, 2023

[Beck et al., 2018] Hylke E. Beck, Niklaus E. Zimmermann, Tim R. McVicar, Noemi Vergopolan, Alexis Berg, and Eric F. Wood. Present and future koppen-geiger climate classification maps at 1-km resolution. Scientific Data, 5(1):180214, Oct 2018

[Cai et al., 2023] Zhiwen Cai, Haodong Wei, Qiong Hu, Wei Zhou, Xinyu Zhang, Wenjie Jin, Ling Wang, Shuxia Yu, Zhen Wang, Baodong Xu, et al. Learning spectral-spatial representations from vhr images for fine-scale crop type mapping: A case study of rice-crayfish field extraction in south china. ISPRS Journal of Photogrammetry and Remote Sensing, 199:28–39, 2023

[Chen et al., 2018] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation, 2018

[Chen et al.

Improvements for AI systems

Bold header: Attentive Feature Injection for Domain Bridging

MineC2FNet employs a teacher feature extractor (a frozen copy of the teacher’s backbone) alongside a student feature extractor (with trainable parameters) operating in parallel, allowing the student to incorporate both generalized source domain features that uses coarse data and specific target domain features that uses fine data.

Bold header: Hybrid Distillation for Selective Knowledge Transfer

The framework implements a hierarchical distillation loss, combining image-level and pixel-level selection via a hybrid mask, ensuring that distillation is applied only when the teacher’s prediction at the pixel level is more accurate than the student’s prediction, which leads to a reduction in False Positive Rate by approximately 29–40% relative to MDIL-SS.

Bold header: Enhanced Boundary Refinement via LBoundary

The model incorporates an explicit boundary loss, defined as a weighted binary-cross entropy loss as described in Eq. (5), where the weights are derived from image edges to emphasize optimizing the boundary regions of the segmented images, which forces the student to pay more attention to correctly classifying pixels on the segmented boundaries.

Bold header: Optimized Coarse Data Curation

The system utilizes a curated coarse dataset, specifically using a version termed CoarseFineFiltered, which is designed to align the coarse data’s characteristics with our fine dataset, using the smallest fine mine’s area (9.81 km2) as a benchmark, ensuring that pre-training on this domain-relevant set yields the highest F1-score of 0.5931 and superior boundary delineation.

Bold header: Robustness Across Climate Zones

The resulting MineC2FNet model demonstrates global robustness and generalizability by performing well across five Koppen–Geiger climate zones, achieving a peak mIoU of 77.34% in temperate climates, proving the model can effectively differentiate mining footprints across diverse geographic locations.

Sources

U-Net: Convolutional Networks for Biomedical Image Segmentation