Multisensor and Multitemporal Dataset of Megathyrsus maxi-mus (Mombasa) across Scales: Tussock-Level Proximal VIS–NIR Spectroradiometry, Thermal, and RGB Imaging with Field-Level UAV Multispectral and Satellite Data under Trop-ical Silvopastoral Conditions across Wet and Dry Seasons

Abstract

Multisensor, Multitemporal Forage Dataset: Megathyrsus maximus cv. Mombasa under Tropical Silvopastoral Conditions (Wet and Dry Seasons, Colombian Caribbean) Overview This dataset comprises 400 structured observations of individual Megathyrsus maximus cv. Mombasa tussocks collected at the AGROSAVIA Turipaná Research Centre (Cereté, Córdoba, Colombia; 8°53′N, 75°49′W) during two contrasting climatic periods: Wet season: August 16 – September 13, 2024 Dry season: March 7 – April 4, 2025 Forty georeferenced tussocks were monitored under two light conditions: 19 shaded tussocks and 21 full-sun tussocks. Measurements were acquired every 5 days at regrowth days 7, 14, 21, 28, and 35, generating 200 observations per season (95 shaded and 105 full-sun observations per season). The dataset integrates proximal sensing, RGB imaging, UAV multispectral imagery, Sentinel-2 satellite products, environmental variables, and destructive laboratory measurements to characterize forage growth dynamics, canopy spectral response, biomass accumulation, and nutritional quality under tropical silvopastoral conditions. Documentation Files This repository includes two documentation files in addition to the master dataset: Data_Dictionary_BD_Forage2Season.xlsx — a complete, column-by-column dictionary of all 107 variables in the Dataset-Wet-Dry sheet, specifying variable name, description, category, data type, unit/format, source/origin, and processing level (Raw vs. Derived, including file-reference columns). It also documents the missing-value code ("N/A") used throughout the dataset. File_Manifest_BD_Forage2Season.xlsx — a manifest listing every file referenced by the tabular dataset (VIS–NIR spectral replicates, RGB and thermal imagery, UAV multispectral orthomosaics, and Sentinel-2 products), organized by category with file counts and relative paths. Dataset Structure The master file BD_Forage2Season.xlsx contains 400 rows × 107 columns, organized into the following variable groups: Identification and Geospatial Information (12 columns)Sampling date, season, regrowth day, tussock location, light condition, tussock identifier, and geographic coordinates in WGS84 format (degrees, minutes, seconds). Morphological and Structural Variables (5 columns)Number of stems, number of leaves, tussock height (cm), measurement height, and canopy angle. VIS–NIR Proximal Spectral Variables (44 columns)Vegetation indices and spectral variables derived from proximal VIS–NIR measurements: NDVI, GNDVI, NDRE, SRI, TVI, SAVI, OSAVI, ARVI, SARVI, SARVI2, EVI2, NLI, VARI, CLGR, CLRE, NDWI, RDVI, WDRVI, NGRVI, GRVI, LAI, ARI1, ARI2, BGI, BRI, NPCI, NPQI, PSRI, and SIPI. The dataset also includes calibrated spectral reflectance values for NIR, Red, Green, Blue, and RedEdge bands, maximum reflectance, logarithmic spectral transformations, and first derivative spectral features. Within this group, the five key-wavelength reflectance values and the spectral maximum constitute raw sensor output, while the log-transformed bands, the red-edge derivative, and the 29 vegetation indices are derived/processed products computed from that raw reflectance data (see Data_Dictionary for the per-column breakdown). Four replicate spectra were collected per tussock (40) per sampling date (10), using a calibrated VIS–NIR spectrometer referenced with a Spectralon diffuse reflectance standard. RGB-Derived Canopy Metrics and Covariance Features (14 columns)Variables extracted from RGB image segmentation using a Mahalanobis distance-based methodology implemented in Python (numpy, scipy; Python 3.12). Reference pixels for the covariance matrix were manually sampled from a single representative leaf per tussock. Includes: green pixel fraction, mean green value, Mahalanobis deviation metrics, Mahalanobis distance, pixel count, greenness classification, Boolean canopy cover indicators, and covariance features derived from R, G, B, RedEdge, and NIR channels. Environmental Microclimate Variables (4 columns)Relative humidity (%), solar radiation, precipitation (mm), and ambient temperature (°C). These variables are recorded once per sampling date and shared across all tussocks measured on that date. Biomass and Nutritional Quality Variables (14 columns)Fresh biomass, dry biomass subsamples, dry matter production, dry matter percentage, and laboratory-derived nutritional parameters: crude protein (CP), crude fiber (CF), neutral detergent fiber (NDF), acid detergent fiber (ADF), ether extract (EE), ash (ASH), and lignin (LIG). Also includes biomass classification categories. Imagery and File Path References (14 columns)Structured file paths linking each observation with associated sensing products: four VIS–NIR spectral replicate files, RGB tussock imagery, thermal imagery, UAV multispectral orthomosaics (Green, Red, RedEdge, NIR, NDVI), and Sentinel-2 VIS and NDVI products (available for only 2 of the 10 sampling dates — see below). See File_Manifest_BD_Forage2Season.xlsx for the complete list of files referenced by these columns. Morphological, spectral, imagery, and biomass (fresh/dry weight) variables are recorded independently for each of the 400 tussock-date observations. The seven forage nutritional-quality parameters (CF, NDF, ADF, LIG, CP, EE, ASH), by contrast, were derived from 40 group-level composite laboratory samples — one per tussock, pooling biomass across its sampling dates — and the same composite value is repeated across the rows corresponding to that tussock. Users should therefore treat the nutritional-quality columns as tussock-level (n = 40), not as 400 independent replicates, when conducting statistical analyses. Environmental/microclimate variables are recorded per sampling date and shared across all tussocks measured on that date (n = 10 independent values). Complementary Files RGB Imagery — High-resolution tussock images acquired using an Apple iPhone 15 (48 MP). Thermal Imagery — Thermal images acquired using a FLIR 300 long-wave infrared sensor (7.5–14 µm). UAV Multispectral Orthomosaics — Radiometrically calibrated (Spectralon reference) GeoTIFF orthomosaics generated from flights performed with a DJI Mavic 3 Multispectral at approximately 50 m altitude, including Green, Red, RedEdge, NIR, and NDVI products. Ten unique UAV flight acquisitions were performed, one per sampling date, covering the 21 full-sun tussocks (imagery is not applicable to shaded tussocks, which are not attributable at the landscape level). VIS–NIR Spectral Data — Forty folders containing replicate spectral measurements stored as .txt files for each tussock and sampling date. Sentinel-2 Imagery — Data Availability and ProcessingSentinel-2 Level-2A products are delivered pre-calibrated radiometrically and atmospherically corrected by the data provider (Sen2Cor L2A processing). Scenes with cloud cover below 30% were evaluated for the study area and the ten field-observation dates. Suitable Sentinel-2 imagery meeting the spatial, temporal, and cloud-coverage criteria was available for only 2 of the 10 field-observation dates (2024-08-16 and 2024-08-23). The Excel dataset identifies the Sentinel-2 image associated with the corresponding observations for these two dates only; the remaining eight dates are coded as "N/A" in the plot-Sentinel2-VIS and plot-Sentinel2-NDVI columns, as no suitable acquisition was available within the defined temporal criteria. The available Sentinel-2 files provided in this dataset correspond to processed PNG visualizations generated from the original Level-2A imagery and are therefore classified as processed products, not raw Sentinel-2 data. UAV imagery is available for 210 of the 400 observations (the 21 full-sun tussocks across all 10 sampling dates). Missing observations were primarily associated with tropical cloud cover and adverse weather conditions, or with tussocks under shade canopy for which landscape-level imagery is not attributable. Missing values within the dataset are coded as "N/A". Key Descriptive Statistics Variable Dry Season (mean ± SD) Wet Season (mean ± SD) Height (cm) 58.3 ± 19.2 104.4 ± 30.4 Fresh biomass (g) 182.5 ± 191.6 469.5 ± 314.6 Dry matter (%) 20.6 ± 2.8 16.4 ± 1.8 Crude protein (%) 15.2 ± 1.7 14.1 ± 3.2 NDVI (VIS–NIR) 0.898 ± 0.088 0.912 ± 0.086 Temperature (°C) 34.9 29.3 Precipitation (mm/day) 4.5 13.5 One-way ANOVA analysis revealed highly significant seasonal differences (p < 0.001) in most evaluated variables, while NDVI showed no statistically significant seasonal variation (p = 0.125). Applications This dataset supports: Machine learning models for non-destructive biomass and forage quality estimation Multivariate analysis integrating proximal and remote sensing data Comparative evaluation of sensing platforms (proximal, UAV, and satellite) Precision grazing management in tropical silvopastoral systems Development of AI-based forage monitoring systems for Agriculture 4.0 applications Related Publication AgriEngineering 7(4):111 — Serpa-Imbett, C.M. et al. (2025). AgriEngineering, 7(4), 111. Note: The associated publication used only the wet-season subset (n = 200) of the present dataset. License CC BY 4.0