Problem and Chronological Development

Problem and Chronological Development

The problem

Precision agriculture promises that GIS, remote sensing, and machine learning can turn sparse field sampling into actionable management maps — but three practical questions are rarely tested together in one study:

1. Does spatial interpolation of soil/water chemistry actually produce usable variable-rate prescription maps for a real field?
2. Can pheromone-trap pest-monitoring data be mapped against terrain and land use to support targeted intervention, rather than blanket spraying?
3. When machine learning is added to spatial interpolation, does it actually improve prediction accuracy for a real agronomic target (grape ripening / Brix) — or is that benefit assumed rather than demonstrated?

This project answers all three, using real field data from three linked University field practicals at UTAD (Vila Real, Portugal), and reports what was found honestly — including where the more sophisticated method did not win.

Chronological development

1. Maize fertility zoning (Coimbra-region field)

Problem: point-sampled soil chemistry and water quality data needed to become continuous, actionable variable-rate fertiliser prescription maps.

What was done:

• Delineated the study area and characterised its terrain (Figure 1)
• Applied inverse-distance-weighted (IDW) interpolation to build continuous surfaces from point samples (Figure 2)
• Tracked NDVI across the growing season to understand crop vigour dynamics independently of the soil sampling (Figure 3)
• Designed and interpolated a water-sampling network (Figure 4)
• Ran descriptive statistics and correlation analysis across all measured variables (Figure 14)
• Built and evaluated zonation models (Figure 15)
• Converted model outputs into variable-rate prescription maps ready for field application (Figure 16)

2. Vineyard pest monitoring (Quinta da Senhora da Graça)

Problem: pheromone-trap moth-catch counts are collected at discrete trap locations — turning that into a map that tells a grower where intervention is actually needed requires linking catch data to terrain and land use.

What was done:

• Characterised the vineyard's terrain (Figure 5)
• Mapped land use and grape varietal composition across the site (Figure 6)
• Built the pheromone-trap network and analysed spatio-temporal pest dynamics across the season (Figure 7)

3. Grape ripening / Brix prediction (Quinta de Nossa Senhora de Lurdes) — the core methodological test

Problem: does adding machine learning and regression-kriging to spatial Brix (sugar-ripeness) prediction actually outperform simple interpolation?

What was done, in order:

1. Characterised the study site (Figure 8)
2. Built an initial regression-based ripening forecast (Figure 9)
3. Interpolated Brix from traditional field sampling (Figure 10) and from a second, independent set of UTAD Enology experimental plots (Figure 11) — a deliberate cross-check using two independent sampling designs
4. Partitioned the site by Thiessen/Voronoi polygons for a third spatial view of the same data (Figure 12)
5. Ran full regression diagnostics alongside remote-sensing products (Figure 13)
6. Trained and cross-validated seven competing models — IDW, Random Forest, Gradient Boosting, a Neural Network (MLP), a Stacking Ensemble, and two hybrid regression-kriging variants (RF-RK, GB-RK) — under spatially grouped cross-validation, holding each sampling location out in full across all five sampling dates (Figure 17)
7. Computed permutation feature importance to see what was actually driving predictions (Figure 18)
8. Computed global Moran's I to test how much spatially structured signal was even present in the raw Brix values (Figure 19)
9. Plotted observed-vs-predicted values out-of-fold, the honest test of real predictive skill (Figure 20)
10. Generated the final spatial interpolation surface for a representative date (30 August) (Figure 21)

The result — reported honestly, not smoothed over

Under this rigorous spatial cross-validation, none of the five ML models or two hybrid regression-kriging variants beat simple IDW interpolation:

Model	Cross-validated R²
IDW	0.538
Random Forest	0.520
Stacking Ensemble	0.467
Gradient Boosting	0.453
Regression-kriging (RF-RK / GB-RK)	0.378–0.454
Neural Network (MLP)	0.370

Why, diagnosed rather than assumed: permutation importance showed Julian day (i.e. simple seasonal timing) dominates NDVI and spatial coordinates by roughly 5:1 — ripening is mostly a time signal, not a space signal at this site. Global Moran's I on raw Brix values was only I = 0.005–0.040 (against an expected −0.015 under complete spatial randomness) — there just wasn't much spatially structured residual signal left for regression-kriging to exploit once the temporal trend was accounted for.

Conclusion carried forward: this isn't a failure of the method — it's a finding about this dataset's sampling density (68 points) and its temporal-vs-spatial signal balance. The maize case study, with denser, multi-covariate sampling, is identified as the leading candidate where the same pipeline would likely show a clearer ML/regression-kriging advantage — this is exactly what the companion projects in this author's other repositories go on to test.

See the Figure Summary page for every figure referenced above, embedded with full captions.