Sm00thix/unet_chemical_map

PorkBellyHSI

This repository contains source code showing the model architecture, loss function, and training pipeline used by Engstrøm et al. [1] to generate chemical maps of pork bellies with a modified U-Net [2]. Executing train_unet_chemmap.py will train, validate, and evaluate the modified U-Net under the five-fold cross-validation scheme explained by Engstrøm et al. [1]. Likewise, load_ensemble_unet.py loads an ensemble of the five U-Nets (weights stored in model_weights/) and uses them to make the predictions shown in ensemble_prediction.png and ensemble_prediction_masked.png.

Note that these scripts are for documentation purposes as actual training and evaluation require access to the dataset by Albano-Gaglio et al. [3].

If you want a U-Net implementation, this repository releases a U-Net implementation under the permissive Apache 2.0 License.

References

1. O.-C. G. Engstrøm, M. Albano-Gaglio, E. S. Dreier, Y. Bouzembrak, M. Font-i-Furnols, P. Mishra, and K. S. Pedersen (2025). Transforming Hyperspectral Images Into Chemical Maps: An End-to-End Deep Learning Approach

2. O. Ronneberger, P. Fischer, and Thomas Brox (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI 2015.

3. M. Albano-Gaglio, P. Mishra, S. W. Erasmus, J. F. Tejeda, A. Brun, B. Marcos, C. Zomeño, and M. Font-i-Furnols (2025). Visible and near-infrared spectral imaging combined with robust regression for predicting firmness, fatness, and compositional properties of fresh pork bellies Meat Science.

Funding

This work has been carried out as part of an industrial Ph.D. project receiving funding from FOSS Analytical A/S and The Innovation Fund Denmark. Grant number 1044-00108B.

The data used to train the models yielding the weights in model_weights/ was collected during a project receiving funding from MICIU/AEI /10.13039/501100011033/ and FEDER ‘Una manera de hacer Europa’ [grant number RTI2018-096993-B-I00, 2019-2022]; and the Spanish National Institute of Agricultural Research (INIA) [grant number PRE2019-089669, 2020-2024].