Deep learning framework for automated otolith growth increment counting in Chloroscombrus chrysurus, a subtropical fish.

Published: 4 August 2026| Version 1 | DOI: 10.17632/j3f8kc64ps.1
Contributors:
,
,
,
,

Description

Reliable age estimation is essential for understanding fish population dynamics and supporting sustainable fisheries management. Traditionally, fish age estimation relies on manual counting of growth increments in otoliths, a procedure that requires extensive expertise, is time-consuming, and may be affected by reader subjectivity. This dataset supports the development and evaluation of a deep learning framework for automated otolith increment counting and age estimation in the Atlantic bumper (Chloroscombrus chrysurus). The main objective of this study was to evaluate the feasibility of a deep learning approach for automating the counting of otolith growth increments in a subtropical fish species. The dataset comprises 998 sagittal otolith images (TIFF format) collected from individuals sampled along the southeastern Brazilian Bight and provided by the Collection of Calcified Fish Structures – FishCAST (Vaz-dos-Santos, 2026). Each image is associated with biological information in a metadata file (CSV format), including total length and manually validated growth increment counts (Nr). The dataset includes five increment classes (Nr-3, Nr-4, Nr-5, Nr-6, and Nr-7), representing the biological categories used for model training and evaluation. The provided Python scripts implement a convolutional neural network based on the InceptionV3 architecture, adapted through transfer learning and multi-task learning to simultaneously perform increment-class classification and fish length prediction. The computational workflow includes image preprocessing, dataset organization, data augmentation, model training and fine-tuning, prediction generation, and model performance evaluation. Results obtained from this dataset demonstrate the feasibility of using deep learning for automated otolith increment classification in a subtropical fish species with complex increment periodicity. The model correctly classified increment classes with an overall accuracy of 55.5%, showing higher predictive performance for older increment classes (Nr-5 to Nr-7) and moderate performance for younger classes (Nr-3 and Nr-4). These predictions can subsequently be interpreted biologically by converting increment classes into age estimates according to validated species-specific increment formation patterns.

Files

Steps to reproduce

1. Environment Setup • Install Python 3.11 or later. • Install the required dependencies: tensorflow (2.20), keras (3.x), pandas, numpy, scikit-learn, matplotlib, and seaborn. •Optional but recommended: Configure a CUDA-enabled GPU to handle the InceptionV3 backbone training efficiently. 2. Data Preparation and Directory Structure The script expects a specific organization in the root directory: • Images: Create a folder named /CHCH_All/ containing the otolith images in .tif format. • Metadata: Provide a file named Biologic_data.csv in the root directory. • The CSV must use a semicolon (;) as a separator. • Columns required: ID (matching the image filenames), Length, and Nr (increment counts). • Note: The script automatically filters the data to keep only Nr classes between 3 and 7. 3. Execution Pipeline Run the main Python script to start the two-phase training process: python your_script_name.py • Phase 1 (Transfer Learning): The model trains the custom classification and regression heads while keeping the InceptionV3 backbone frozen. • Phase 2 (Fine-Tuning): The last 15 convolutional layers of the backbone are unfrozen and trained with a lower learning rate to optimize feature extraction for otolith patterns. 4. Outputs and Results Upon completion, the script generates several files in the root directory: • Model: Classification_CHCH_Multitask.keras (the final trained weights). • Metrics: classification_report.csv, confusion_matrix.csv, and regression_metrics.json. • Predictions: predictions_complete.csv (contains original vs. predicted values for the test set). • Visualizations: training_metrics_publication.png and .pdf (high-resolution accuracy and loss curves).

Institutions

Categories

Automation Process, Fish Growth, Fish Age Structure, Convolutional Neural Network, Deep Learning, Inception V3

Funders

Licence