Moya, AliciaInés, AdriánCedrón, FranciscoFernández-Lozano, CarlosFernández-Martínez, AlejandroHeras, Jónathan2026-03-092026-03-092026-02A. Moya, A. Inés, F. Cedrón, C. Fernández-Lozano, A. Fernández-Martínez, and J. Heras, "Comparative evaluation of deep learning architectures for bioclast classification in atomic force microscopy images", Systems and Soft Computing, Vol. 8, June 2026, 200458, https://doi.org/10.1016/j.sasc.2026.2004582772-9419https://hdl.handle.net/2183/47625The code needed to train each of the models mentioned in the paper can be found at https://github.com/adines/BioclastClassification. This repository includes a Python file for training the base models with different sizes, another Python file for training the models using the progressive resizing technique, and a final Python file for creating the ensembles. In addition, all data along with the splits are available at https://doi.org/10.5281/zenodo.15689362.[Abstract]: The identification and classification of bioclasts in limestone are fundamental tasks in petrographic analysis, traditionally performed through optical microscopy and expert-driven interpretation. While Atomic Force Microscopy (AFM) provides high-resolution topographic information at micro- to nanoscales, the analysis of AFM images for bioclast identification remains challenging due to complex surface morphologies, multi-scale textures, and the lack of direct correspondence with optical observations. Existing approaches rely heavily on manual inspection and complementary imaging techniques, resulting in time-consuming and non-scalable workflows. In this work, we address the problem of automated bioclast classification directly from AFM topographic images by systematically evaluating eight state-of-the-art deep learning architectures under multiple input resolutions and training strategies. In particular, we investigate the impact of image resolution, progressive resizing, and multi-scale feature modeling on classification performance. Our comparative analysis reveals a strong positive correlation between input resolution and performance, with progressive resizing consistently improving model robustness. Among the evaluated architectures, HRNet-based models demonstrate superior performance in capturing hierarchical geological textures, achieving a maximum F1-score of 79.1. Furthermore, an ensemble of the three best-performing HRNet variants further enhances classification accuracy, reaching an F1-score of 82.1.engAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Atomic force microscopyBioclast classificationTexture analysisDeep learningArtificial neural networksjournal articleopen access10.1016/j.sasc.2026.200458