Dataset for predicting maize growth, nitrogen uptake, and yield using UAV remote sensing and machine learning.

Dataset: All_variables.xlsxDescription: This dataset contains field-measured crop phenological variables and UAV-derived spectral reflectance bands and vegetation indices used for modeling crop growth, nitrogen status, and yield-related traits. Variables include:- Phenological variables: LAI (leaf area index), SPAD (chlorophyll content), PH (plant height), AGB (aboveground biomass), N_Uptake- Spectral reflectance bands: Blue, Green, Red, Red-edge, NIR- Vegetation indices: NDVI, NDRE, GNDVI, SR, SRre, CIre, CIg, EVI Units:- LAI: m² m⁻²- SPAD: unitless (SPAD meter reading)- PH: cm- AGB: Mg ha⁻¹- N_Uptake: kg ha⁻¹- VIs: unitless Data source:- UAV multispectral imagery and field measurements collected during maize growing seasons (2021–2022) near College Station, TX. Processing:- Data were cleaned by removing missing values- Numeric conversion applied where necessary- Used for machine learning model training and validation Missing values:- Represented as NA and removed prior to analysis File: RF.R Description:This script implements Random Forest (ranger) models to predict crop phenological variables (LAI, SPAD, AGB, N uptake, plant height) using UAV-derived spectral data. Key steps:- Data import from Excel file- Data cleaning and preprocessing (NA removal, numeric conversion)- Train/test split (80/20)- Hyperparameter tuning using repeated cross-validation (5x5)- Model training using ranger- Model evaluation (R², RMSE, MAE)- SHAP value computation and visualization- Variable importance analysis Inputs:- Reflectances.xlsx Outputs:- Model performance metrics- SHAP plots- Observed vs predicted plots- Variable importance plots Dependencies:- readxl, dplyr, caret, ranger, ggplot2, fastshap, shapviz, patchwork Reproducibility:- Random seed set to 123 File: SVM.R Description:Implements Support Vector Machine (radial kernel) models using caret. Key features:- Hyperparameter tuning (C and sigma)- Data standardization (center/scale)- Cross-validation (5x5)- Performance evaluation and plotting File: XGBoost.R Description:Implements gradient boosting models (XGBoost) for predicting crop traits. Key features:- Hyperparameter tuning using caret- Feature importance (gain)- SHAP value computation- Export of predictions and plots File: SMLR.R Description:Implements stepwise multiple linear regression using AIC. Key features:- Variable selection using stepAIC- Multicollinearity control using VIF- Cross-validation (manual 5x5)- Model diagnostics and coefficient export File: Fig1.pngDescription: (A) Map showing the study location in Texas, USA; (B) UAV-derived multispectral image of experimental maize plots. Purpose:- Provides geographic context and experimental layout Data source:- Field experiment near College Station, TX- UAV multispectral imagery File: Correlation_matrix.pngDescription: Pearson correlation matrix showing relationships among phenological variables, spectral reflectance bands, and vegetation indices. Key variables:- Phenological: LAI, SPAD, PH, AGB, N uptake- Spectral: Blue, Green, Red, Red-edge, NIR- Vegetation indices: NDVI, NDRE, GNDVI, SR, SRre, CIre, CIg, EVI Interpretation:- Positive correlations observed between phenological traits and NIR/vegetation indices- Negative correlations with visible bands Purpose:- Used to guide feature selection and modeling strategy File: AGB.PNGDescription: Observed vs predicted aboveground biomass (AGB) using four models:A) Stepwise Multiple Linear Regression (SMLR)B) Support Vector Machine (SVM)C) Random Forest (RF)D) XGBoost Details:- Green points: training data- Orange points: testing data- Black line: 1:1 reference line Performance:- RF and XGBoost showed the highest accuracy (R² up to 0.96–0.99) Purpose:- Compare model performance for biomass prediction File: LAI.PNGDescription: Observed vs predicted leaf area index (LAI) across four models (SMLR, SVM, RF, XGBoost). Purpose:- Evaluate model accuracy for canopy structure estimation Key result:- RF and XGBoost outperform SMLR File: N Uptake.PNGDescription: Observed vs predicted plant nitrogen uptake using four models. Key insight:- Machine learning models (RF, SVM, XGBoost) outperform linear regression File: Plant Height.PNGDescription: Observed vs predicted plant height (cm) across four models. Key result:- XGBoost and RF show the strongest predictive performance File: SPAD.PNGDescription: Observed vs predicted leaf chlorophyll content (SPAD). Observation:- Lower prediction accuracy compared to other traits, especially for SMLR File: SHAP_beeswarm_grid_all_5.pngDescription: SHAP (SHapley Additive exPlanations) beeswarm plots showing feature importance and contribution for five response variables:A) LAIB) SPADC) Nitrogen uptakeD) AGBE) Plant height Details:- X-axis: SHAP value (impact on model output)- Color: feature value (low to high) Purpose:- Interpret model predictions and identify key spectral drivers Software:- R (fastshap, shapviz, ggplot2)