Plot-level VNIR–SWIR hyperspectral reflectance dataset for potato under nitrogen fertilization and field-to-laboratory measurement conditions
Description
This dataset contains 180 plot-level averaged visible–near-infrared to short-wave infrared (VNIR–SWIR) reflectance spectra from Russet Burbank potato grown under nitrogen fertilization treatments near Carberry, Manitoba, Canada, during the 2024 growing season. A randomized complete block design comprised 15 nitrogen treatments and four replicates, giving 60 experimental plots. Each plot is represented by field-canopy, laboratory fresh-leaflet and laboratory dried-powder spectra, linked to petiole nitrate-N (PNN) measurements and treatment metadata. Spectra were measured using an ASD FieldSpec Pro spectroradiometer and exported over 350–2500 nm on an interpolated 1-nm wavelength grid. This spacing does not represent the instrument’s native spectral resolution. The lowest-level deposited spectral products are plot-level averaged reflectance spectra before dataset-level cleaning; individual acquisitions are not included. The repository provides averaged spectral exports, long- and wide-format tables, cleaned reflectance spectra, preprocessed spectra, quality-control summaries, method documentation and reconstruction scripts. Cleaned spectra exclude the inclusive wavelength intervals 350–400, 1290–1500, 1790–2050 and 2350–2500 nm. Preprocessing follows Savitzky–Golay smoothing, standard normal variate correction and a Savitzky–Golay first derivative. Both Savitzky–Golay operations use an 11-point window and polynomial order 2; complete numerical settings are documented in the repository. PNN was determined by AGVISE Laboratories using 2% acetic-acid extraction and cadmium-reduction flow-injection colorimetry. Results were reported in ppm on a dry-matter basis and converted to percent by dividing by 10,000. The dataset supports crop-spectroscopy and laboratory-to-field calibration-transfer investigations within the represented experimental conditions. Transfer to other cultivars, sites, seasons, growth stages or instruments requires independent validation. The dataset is associated with the research article “VNIR–SWIR hyperspectral spectroscopy for nitrogen assessment in potato crops: Deep learning and regression models across field and laboratory conditions,” published in Smart Agricultural Technology.
Files
Steps to reproduce
Extract the dataset ZIP and open README.md, metadata/variable_dictionary.csv and metadata/preprocessing_protocol.md. Plot-level averaged reflectance spectra before dataset-level cleaning are provided in plot_level_averaged_spectra/. Individual spectral acquisitions are not included. To reproduce the deposited spectral tables, install the dependencies listed in scripts/requirements.txt and run scripts/rebuild_averaged_tables.py from the main dataset folder. The script reconstructs all 20 spectral tables, compares them with the deposited tables and writes outputs to regenerated/ without overwriting the deposited files. Long-format and wide-format tables are available in processed_data/long_format/ and processed_data/wide_format/. Cleaned reflectance spectra are available in processed_data/clean_reflectance_noisy_regions_removed/. Cleaning removes the inclusive intervals 350–400, 1290–1500, 1790–2050 and 2350–2500 nm, retaining 1,477 wavelength columns. Preprocessed spectra are available in processed_data/preprocessed_SG_SNV_FD/. The operation sequence is wavelength removal, Savitzky–Golay smoothing, standard normal variate correction (SNV), and a Savitzky–Golay first derivative. Both Savitzky–Golay operations use an 11-point window, polynomial order 2 and interpolation boundary handling. Derivative orders are 0 for smoothing and 1 for the first derivative, with delta = 1.0. SNV subtracts each smoothed spectrum’s mean and divides by its population standard deviation (ddof = 0). Filtering is applied across the concatenated retained wavelength columns; windows therefore cross the joins between retained intervals. Treatment information, PNN values, plot assignments and quality-control summaries are provided in metadata/. The confirmed PNN analytical method is documented in metadata/pnn_analytical_method.md, optical geometry in metadata/optical_geometry.md, and unresolved source-record differences in metadata/record_status.md. All 60 plot-level spectra per measurement condition were retained.
Institutions
- University of ManitobaManitoba, Winnipeg