# A hybrid deep boosting framework with adaptive label stabilization for SEM-based porosity estimation in fly-ash cement mortar

## TL;DR

A new deep learning framework improves automatic porosity estimation in cement materials using SEM images, reducing manual work and increasing accuracy.

## Contribution

A hybrid deep-boosting framework with adaptive label stabilization and fusion of semantic and texture features for robust porosity estimation.

## Key findings

- The framework achieves high accuracy (R2 = 0.9816) in porosity estimation from SEM images.
- It outperforms traditional CNN and handcrafted feature methods in robustness and generalization.
- The model provides transparent explanations and works across different imaging conditions.

## Abstract

Accurate measurements of porosity of cementitious matrices are critical in pre- Q7 dicting mechanical behavior, durability, and transportation processes. Traditional methods based on SEM, such as manual thresholding, a simple binarization method, and end-to-end convolutional neural network (CNN) regressors, are, however, highly affected by image contrast variation, polishing quality, magnification, and operator bias. To address these limitations, the current article develops a hybrid deep-boosting framework for fully automatic porosity estimation directly from raw backscattered-electron SEM images of fly-ash cement mortar. The key novelty of the proposed approach lies in the adaptive stabilization of porosity labels and the hybrid fusion of deep semantic and handcrafted texture features, which together improve robustness to imaging artifacts, boundary ambiguity, and overfitting.

Annotation ground-truth porosity is optimized using an Adaptive Porosity Label Stabilizer (APLS) that successively improves Otsu threshold masks, first using entropy measures and morphological consistency measures to reduce label noise. Multiscale semantic representations are learned on a ResNet-18 backbone, which is trained with SimCLR on SEM data, while local statistical texture is captured using handcrafted gray-level cooccurrence map (GLCM) features. The resulting mismatched set of features is combined with a learnable Hybrid Feature Refinement Block (HFRB) together with a Feature-Interaction Attention (FIA) block, which explicitly characterizes inter-scale relationships among convolutional channels and texture regressors. The latent representation is then condensed and regressed using a weighted ensemble including CatBoost, XGBoost, and LightGBM learners.

The proposed methodology achieves R2 = 0.9816, RMSE = 0.0236, and MAE = 0.00875 on a rigorously held out test set, outperforming baseline methods that rely exclusively on CNN features, handcrafted descriptors, or naïve hybrid combinations. The validity, stability, and physical plausibility of the model are ensured through a comprehensive assessment, including ablation studies, domain-shift experiments, uncertainty and stability calibration, and a hybrid explainability framework (Grad-CAM++, SHAP). The architecture does not require any manual segmentation, generalizes across magnifications and imaging conditions, and provides transparent, domain-consistent explanatory visualizations. Overall, the proposed framework represents an important step toward fast, reliable, and scalable SEMbased porosity estimation in cementitious systems.