Dataset for Physiological-Driven Irrigation Scheduling in Ananas comosus via Hybrid Machine Learning (PIML-GB)

Description

Dataset Description This dataset provides a comprehensive multimodal matrix containing N=150 independent observations. It is specifically designed to validate a Physics-Informed Machine Learning (PIML) framework for precision irrigation in 5 hectares of commercial pineapple crops (Ananas comosus var. MD2) located in Tauramena, Colombia. The data covers a six-month vegetative cycle. The repository includes high-resolution multispectral imagery (0.82 cm/px) obtained from UAV flights, microclimatic records from a modular IoT agrometeorological station recorded at 10-minute intervals, and mechanistic soil water balance simulations derived from the FAO-56 Penman-Monteith standard. These resources enable the reproduction of the Gradient Boosting (XGBoost) architecture, which achieved a predictive performance of R^2=0.851 and an overall accuracy of 91.1% in hydric status classification.

Steps to reproduce

Steps to Reproduce To replicate the results of the PIML-GB framework for pineapple irrigation scheduling, follow these sequential steps: 1. Radiometric and Spectral Processing Apply radiometric calibration to the raw multispectral TIF images using the provided MicaSense RedEdge-M conversion model and Calibrated Reflectance Panel (CRP) values. Execute the band alignment script using a 3×3 homographic transformation matrix (Green band as master) to ensure sub-pixel registration. Calculate the 17 Vegetation Indices (VIs) defined in the study (Table 5), ensuring the use of an L=0.16 factor for OSAVI to suppress soil background noise. 2. Mechanistic Water Balance (FAO-56) Open the provided CROPWAT 8.0 project files. Configure the biophysical crop parameters for Ananas comosus (Initial Kc=0.50, Mid Kc=0.30) and soil profile characteristics (TAW 140 mm/m, Zr 0.6 m). Run the simulation using the IoT microclimatic logs to extract the daily soil moisture depletion (SMDepl). 3. Multimodal Data Fusion Synchronize the extracted spectral zonal statistics (μ, σ) from the 25 georeferenced sampling units with the meteorological data and ground-truth soil sensor records (N=150 matrix). Integrate SMDepl and VPD as mechanistic features in the final training dataset. 4. Machine Learning Modeling (PIML-GB) Use the Python scripts provided to train the XGBoost regressor and classifier. Apply the optimized hyperparameters: 200 estimators, maximum depth of 5, learning rate of 0.1, and L1/L2 regularization (α=0.1, λ=1.0). Implement an 80/20 stratified hold-out split for validation. 5. Performance Evaluation Validate predictions against the thermogravimetrically-calibrated ground truth to confirm the R²=0.851 and the DSS classification accuracy of 91.1%.

Institutions

• Universidad de Antioquia

  Antioquia, Medellín

Categories

Licence