Author: Naziru Halilu
This repository accompanies the manuscript "Integrated GIS, Remote-Sensing, and Machine-Learning Precision Agriculture Across Three Portuguese Cropping Systems: Maize Fertility Zoning, Vineyard Pest Monitoring, and Machine-Learning-Enhanced Grape Ripening Prediction."
Workflow sketch
View interactive graphical abstract →
Problem. Farm-scale precision-agriculture case studies are common in training and consultancy settings but are rarely synthesised into a single comparative analysis across contrasting cropping systems.
Methodology. Three independent GIS and remote-sensing field campaigns in northern and central Portugal were integrated into one comparative framework: (i) spatial fertility mapping and variable-rate prescription for a maize field in Coimbra (16.85 ha); (ii) GIS-based monitoring of the European grapevine moth across three generations at Quinta da Senhora da Graça, Douro (42.97 ha); and (iii) Brix-based ripening and harvest-date prediction integrated with NDVI monitoring at Quinta de Nossa Senhora de Lurdes, Vila Real (6 ha). Inverse Distance Weighting, Thiessen/Voronoi tessellation, and multispectral vegetation indices converted point-sampled measurements into continuous management surfaces.
Results. In the maize system, yield correlated strongly with deep soil-water content (r = 0.972) and potassium oxide (r = 0.751); a quadratic yield–fertiliser regression (R² = 0.995) supported a variable-rate NPK/liming programme that cut total lime demand by 2.69 t versus uniform application. In the vineyard pest system, pheromone-trap captures showed a consistent three-generation spatial pattern, with a second-generation surge (trap counts 16–18) driving the principal within-season damage peak. In the ripening system, a logarithmic Brix–Julian-day regression (R² = 0.978) forecast harvest maturity (25.5–27 °Bx) around 13 September, and an NDVI–Brix regression explained 88% of variance (R² = 0.880).
- Figures
- Repository structure
- How to Run the Code
- Results summary
- Data provenance
- License
- Citation
- Related work
All 21 manuscript figures, extracted directly from the manuscript and presented in their original figure order:
Figure 1, Study-area delineation for the Coimbra maize field.
Figure 2, Inverse-distance-weighted interpolation surfaces.
Figure 3, Temporal NDVI dynamics across the growing season.
Figure 4, Water sampling design and interpolated surfaces.
Figure 5, Terrain characterisation of the study vineyard.
Figure 6, Land-use and varietal composition.
Figure 7, Pheromone trap network and spatio-temporal pest dynamics.
Figure 8, Study-site characterisation.
Figure 9, Regression-based ripening forecast.
Figure 10, Spatially interpolated Brix values from traditional field
sampling.
Figure 11, Spatially interpolated Brix values from the UTAD Enology
experimental plots.
Figure 12, Brix values partitioned by Thiessen/Voronoi polygons.
Figure 13, Regression diagnostics and remote-sensing products for ripening
prediction.
Figure 14, Descriptive and correlation analysis for the Coimbra maize
field.
Figure 15, Model performance and zonation outputs.
Figure 16, Variable-rate prescription maps.
Figure 17, Cross-validated model comparison for Brix prediction across
seven models.
Figure 18, Feature importance for Brix prediction.
Figure 19, Global Moran's I spatial autocorrelation of raw Brix values.
Figure 20, Out-of-fold observed-versus-predicted Brix values.
Figure 21, Spatial interpolation surfaces for 30 August Brix.
manuscript/ The final manuscript (.docx, 30 pages), with all in-text citations
hyperlinked to bookmarked reference-list entries (28 references),
20 numbered equations covering IDW, NDVI, Pearson correlation,
quadratic regression, Random Forest, Gradient Boosting, a
Multi-Layer Perceptron, a Stacking Ensemble, regression-kriging
(variogram + kriging system), Moran's I, and RMSE/MAE/R², plus a
Nomenclature table and 21 figures.
code/ All Python source code used in this project:
- train_regression_kriging.py : trains and cross-validates IDW,
Random Forest, Gradient Boosting, a Neural Network (MLP), a
Stacking Ensemble, and hybrid regression-kriging (RF-RK,
GB-RK) models for Brix prediction; also computes Moran's I
and permutation importance.
- make_ml_figures.py : generates the model-comparison,
feature-importance, Moran's I, and observed-vs-predicted
analysis charts.
- make_hybrid_map.py : generates the spatial hybrid
regression-kriging surface map.
- make_grid.py : builds the labelled A/B/C...
figure grids from the field-photo/GIS-map source images.
- translate_legend.py : translates the Portuguese GIS
legend text in the vineyard-pest trap figures into English.
data/ The field data used to train the machine-learning models:
- BRIX_AMT.csv : 68 georeferenced (UTM) sampling points,
Brix at 5 dates (15 Jul – 30 Aug).
- Sample_brix_ndvi.xlsx : same points with paired NDVI values.
- BRIX_AMT_V2.xlsx : original source workbook.
results/ Model outputs:
- merged_brix_ndvi_utm.csv : the cleaned, merged analysis
dataset.
- cv_pooled_spatiotemporal.json : cross-validated RMSE/MAE/R²
for IDW, Random Forest, Gradient Boosting, Neural Network,
Stacking Ensemble, RF-RK, and GB-RK (pooled spatio-temporal,
GroupKFold-by-location design; corresponds to Table 3 in the
manuscript).
- cv_model_comparison.json : per-date, spatial-only
cross-validation results.
- morans_i_brix.json : global Moran's I
spatial-autocorrelation statistic for Brix at each sampling
date (Figure 19).
- permutation_importance.json : permutation importance for
Easting, Northing, Julian day, and NDVI (Figure 18).
figures/ All 21 manuscript figures, extracted directly from the
manuscript, in original figure order (Figure_01–Figure_21).
git clone https://github.com/halilunaziru73-creator/Operationalizing-GIS-and-Machine-Learning-across-Contrasting-Cropping-Systems.git
cd Operationalizing-GIS-and-Machine-Learning-across-Contrasting-Cropping-SystemsRequirements: Python 3, pandas, numpy, scipy, scikit-learn, matplotlib.
scikit-learn's GradientBoostingRegressor is used to implement the
gradient-boosted decision-tree models.
pip install pandas numpy scipy scikit-learn matplotlibcd code
python3 train_regression_kriging.py # trains models, cross-validates, writes results/
python3 make_ml_figures.py # writes machine-learning analysis charts
python3 make_hybrid_map.py # writes the spatial hybrid surface mapUnder spatially grouped cross-validation (each vineyard sampling location held out in full, across all five dates), Random Forest, Gradient Boosting, a Neural Network (MLP), a Stacking Ensemble, and hybrid regression-kriging do not outperform simple IDW interpolation for Brix prediction at the sampling density available (68 points): cross-validated R² is 0.538 (IDW), 0.520 (Random Forest), 0.453 (Gradient Boosting), 0.370 (Neural Network), 0.467 (Stacking Ensemble), and 0.378–0.454 (regression-kriging variants). Two diagnostics explain this result: permutation importance shows Julian day dominates NDVI and spatial coordinates by roughly 5:1, and the global Moran's I statistic shows only weak spatial autocorrelation in raw Brix values (I = 0.005–0.040 against an expected −0.015 under complete spatial randomness), indicating limited spatially structured residual signal for regression-kriging to exploit once the temporal trend is removed. The pipeline is a reusable methodological contribution expected to show clearer benefits on denser, multi-covariate, and/or multi-season datasets, with the maize case study identified as the leading candidate for this extension (see manuscript Sections 4.1 and 4.3).
All data in data/ were collected from original field campaigns conducted as
part of UTAD coursework (Quinta de Nossa Senhora de Lurdes, Vila Real,
Portugal). All cross-validation results in results/ were computed directly
from this dataset.
Released under the MIT License.
If you use this repository, please cite it using the metadata in
CITATION.cff (GitHub renders a "Cite this repository"
button on the repo's main page, in the top-right "About" panel).
Part of a broader body of research on GIS, remote sensing, and machine learning for agronomic and environmental applications: