Published August 19, 2026
| Version v010
Dataset
Open
Hyperspectral Imaging Dataset for Moisture-Content Prediction during Cocoyam (Colocasia esculenta (L.) Schott) Drying
Authors/Creators
Description
1.0 Overview
This dataset supports a machine-learning study of cocoyam drying. For each sample it provides a hyperspectral image (HSI) cube, the process metadata, and the reference moisture content. The intended model is two-step: from the HSI cube and metadata at steps t-1 and t, it estimates the moisture content at the next step, t+1.
The data originate from a hot-air drying study of purple-speckled cocoyam (*Colocasia esculenta* (L.) Schott) slices (see *Related publications* below) and have been reorganised for supervised learning with moisture content as the prediction target.
2.0 Related publications
Please cite the source study for data provenance:
- Ndisya, J.; Gitau, A.; Mbuge, D.; Arefi, A.; Bădulescu, L.; Pawelzik, E.; Hensel, O.; Sturm, B. (2021). Vis-NIR Hyperspectral Imaging for Online Quality Evaluation during Food Processing: A Case Study of Hot Air Drying of Purple-Speckled Cocoyam (Colocasia esculenta (L.) Schott). Processes, 9 (10), 1804. https://doi.org/10.3390/pr9101804
- Companion manuscript: Babor, M., Dey, R., Alsatouf, A., Ndisya, J., Arefi, A., Sturm, B., & Höhne, M. M.-C. (n.d.). Explainable AI-Driven Sparse Hyperspectral Modeling for Moisture Forecasting During Cocoyam Drying. (manuscript under review).
3.0 Dataset contents
| File | Description |
| `hsi.zip` | One MATLAB `.mat` hyperspectral cube per sample, named `{Batch}.{timestep}{slice}.mat` |
| `meta_data.csv` | Per-sample process metadata (1,056 samples) |
| `y.csv` | Prediction target: moisture content (%) |
| `wavelenghts.csv` | 112 near-infrared wavelengths (nm) |
4.0 File descriptions
4.1 `hsi.zip` - hyperspectral cubes
One MATLAB `.mat` file per sample, named `{Batch}.{timestep}{slice}.mat` (e.g. `1.11.mat`). Each file stores a near-infrared hyperspectral cube acquired with the Specim FX17 camera (937 - 1700 nm) described under *Instrumentation*, with the 112 spectral bands listed in `wavelenghts.csv`.
4.2 `meta_data.csv` - process metadata
One row per sample, ordered by `Batch`, then `Time step`, then `Slice ID`.
| Column | Description |
| `sample_id` | Unique sample identifier, written `{Batch}.{timestep}{slice}` |
| `Reference` | `1` if the sample has a measured reference moisture value; `0` if it has none and the value was imputed. |
| `Batch` | Drying-batch number (1 - 16) |
| `Time step` | Step index within the batch (1 - 11) |
| `Slice ID` | Minutes since start of drying (min) |
| `Drying Time` | Spatial slice number (1 - 6) |
| `Drying Temp` | Drying-air temperature (°C) |
| `Air Velocity` | Drying-air velocity (m/s) |
4.3 `y.csv` - moisture content
| Column | Description |
| `sample_id` | Unique sample identifier (matches `meta_data.csv`) |
| `Moisture Content` | Sample moisture content (%) |
4.3 `wavelenghts.csv` - spectral wavelengths
The spectral wavelength axis for the hyperspectral cubes: 112 near-infrared wavelengths spanning 937 - 1718 nm (the FX17 range).
5.0 Sample identification
`sample_id` is written `{Batch}.{timestep}{slice}`. `Batch` is one of 16 drying batches, each under distinct drying conditions. Within a batch, `timestep` runs 1 - 11 (drying time) and `slice` is 1 - 6 (a spatial slice). Each batch therefore has 11 × 6 = 66 samples, giving 1,056 samples in total. Rows are ordered by batch, then timestep, then slice.
6.0 Intended use
Read a batch and slice in time order together with the moisture content for time-series modelling. The reference modelling task predicts moisture content at step *t+1* from the HSI cube and metadata at steps t−1 and t.
Evaluation:
Compute all accuracy metrics (RMSE, R², RPD) only over samples with `Reference = 1`. Samples with `Reference = 0` have imputed moisture values and must be excluded from reported metrics (e.g. RMSE).
7.0 Instrumentation
The hyperspectral cubes were acquired with a single near-infrared imaging system:
- Camera: FX17 (Specim Spectral Imaging Ltd., Oulu, Finland)
- Lens: 17.5 mm OLET 17.5 N f/2.1 (Specim Spectral Imaging Ltd., Oulu, Finland)
- Stage: motorised linear translation stage (Specim Spectral Imaging Ltd., Oulu, Finland)
- Spectral range: 937 - 1700 nm
- Sampling rate: 3.5 nm
The source study also employed a visible-range V10E PFD camera (400 - 1010 nm); those visible bands are not included in this dataset.
8.0 Data collection
Purple-speckled cocoyam slices (4 mm thick, 25 mm diameter) were hot-air dried in an HT-mini convective dryer at three temperatures (40, 60 and 80 °C) and three air velocities (2.0 ± 0.2, 3.0 ± 0.4 and 4.6 ± 0.7 m·s⁻¹). Samples were drawn every 30 min until a final moisture content of approximately 0.11 kgW/kgDM was reached. During imaging, a white reference tile and a dark reference (shutter closed) were captured and used to correct for non-uniform illumination and sensor noise. Full methodological details are provided in the source study.
9.0 Data structure and loading
Each `*.mat` file stores a hyperspectral cube - a three-dimensional array (spatial rows × spatial columns × spectral bands) with an associated wavelength vector. Load the data with MATLAB or Python:
```matlab% MATLABS = load('1.11.mat');% inspect S (e.g. whos -file 1.11.mat) to identify the cube and wavelength variables``````python# Pythonimport scipy.io as siomat = sio.loadmat('1.11.mat')```Note: the exact variable names inside the `.mat` files should be verified against an actual file before scripting your analysis.
10.0 Funding and acknowledgements
This work was supported by the following organisations as recorded in the source studies.:
- Deutscher Akademischer Austauschdienst (DAAD, doctoral research grant no. 91692333)
- Bundesministerium für Ernährung und Landwirtschaft (BMEL, grant no. 323-06.01-03-2816PROC01) [partially]
- Deutsche Forschungsgemeinschaft (DFG) [partially]
Files
hsi.zip
Additional details
Dates
- Available
-
2026-08-26