Limestone thin sections get a lighter backbone: 2.095M parameters, 91.11% accuracy across six classes
Synopsis
Limestone thin-section classification now runs on a compressed MobileNetV3 backbone that folds texture enhancement and multi-level feature fusion into one lightweight network. Fine-grained carbonate grain boundaries and micritic matrix were previously attenuated by successive downsampling, while inter-class similarity stayed high and manual petrographic reading depended on expert judgment. On 3,600 images of six classes collected in Ordos, Xinjiang, Tibet and the Lesser Khingan Mountains, the method reaches 91.11% accuracy with macro precision, recall and F1 of 91.28%, 91.11% and 91.12%, at 2.095M parameters.
Interpretation
Limestone thin-section classification no longer relies on the single-path downsampling of standard MobileNetV3; grain boundaries and texture are enhanced separately at several channel stages before classification. The original MobileNetV3 propagates features through a predominantly single-path topology, and successive downsampling irreversibly degrades delicate textural cues such as grain boundaries, micritic matrix, pores and cements, aggravating inter-class confusion. On the six-class limestone validation set, correct classifications rise from 567 for the baseline to 656, misclassifications fall from 153 to 64, and accuracy climbs from 78.75% to 91.11%.
The network compresses its backbone with simplified inverted residual blocks and spends the freed capacity on a texture enhancement module and a multi-source feature complementation module. Standard MobileNetV3-Large stacks 15 inverted residual blocks with 8 SE attention modules, carrying excessive parameters and computation, and its layer-by-layer representation lacks explicit cross-scale fusion. The simplified inverted residual stack uses 6 modules, 0.332M parameters and 0.510G FLOPs, which are 11.80% and 43.38% of the standard inverted residual stack.
The texture enhancement module extracts information through parallel edge and texture branches, then reweights channels with a full-channel mapping attention. Inverted residual blocks aim at lightweight feature representation and lack targeted enhancement of rock edge and texture information, leaving fine-grained discriminative features underrepresented. The module is embedded at five stages with 24, 40, 80, 112 and 160 channels, and the full-channel mapping attention costs about 0.047M parameters, 2.23% of the 2.095M model.
The multi-source feature complementation module aligns shallow texture, intermediate structure and deep semantics to a common scale before gated fusion. Using only the highest-level feature discards shallow detail, while direct cross-stage concatenation suffers from spatial and channel misalignment and introduces redundancy. In ablation, adding this module raises accuracy from 88.75% to 89.03%, and the complete pipeline reaches 91.11%.
Perspective
The result rests on six classes of limestone thin-section images collected in Ordos, Xinjiang, Tibet and the Lesser Khingan Mountains; 1,286 original images were expanded by region cropping to 600 per class, 3,600 in total, split 8:2 into 2,880 training and 720 validation images. Evaluation covers mudstone, wackestone, packstone, grainstone, intraclast limestone and floatstone, with macro-averaged metrics. The method addresses research and engineering settings that need automatic carbonate thin-section classification on a lightweight model.
Single-model validation accuracy across three random seeds ranges from 88.19% to 89.44%, and the final 91.11% comes from a weighted three-model ensemble with multi-scale test-time augmentation. Residual errors concentrate among mudstone, wackestone and packstone, reflecting gradual transitions in micritic matrix content and grain-support relationships. Data scale, patch-level partitioning and the deployment overhead of multi-model multi-view inference are listed as current limitations.