Choose how to read
The full view retains the technical content, figures and sources. The plain-language explanation introduces the ideas without changing the findings.
Checking a model against a simple water-storage calculation
What this dashboard shows
This mock-up compares trained neural models with a simple water-storage simulator driven by historical rainfall. The target values come from that simulator, not from measured soil moisture or crop yields. It is a separate extension of the original paper’s experiment.
How the model was chosen
Eight settings were run with three different training seeds, giving 24 saved checkpoints. The settings were selected using validation error. The selected models were then checked on separate historical years.
What the adjustment does
A mathematical projection moves predicted water amounts into allowed ranges and enforces a total-water constraint. Because the simulator’s target is inside that allowed set, this adjustment cannot increase the total one-step squared error. That mathematical fact is not evidence that the model learned crop physics.
Where it still fails
Average storage error improved, but average water-use error worsened. Some season-long runs also became worse after adjustment. A model can obey the water constraints and still predict the wrong amounts. The simple exact simulator remains the appropriate engine here; no speed advantage or farm benefit has been shown.
Use the full view for the complete technical discussion, figures, interactive elements and sources.
Physical consistency is not a crop forecast
Actual historical rainfall drives synthetic soil-water experiments near Pedro Plains, Jamaica. These results test a numerical emulator, not field yield, income or future rainfall.
Three different uses
Forecast: not tested here. A future-rainfall forecast needs issue-time data and independent forecast verification.
Impact scenario: assumed soil and economics under historical rainfall; not a prediction of farm profit.
Surrogate diagnostic: neural output compared with the exact water-balance bucket; targets are synthetic.
What changed
Validation chose physical penalty 0.01 and 3,000 updates. A separate exact feasibility projection reduces mean one-step error from 1.644 to 1.163 mm and average seasonal storage error from 17.54 to 6.45 mm. Water-use error worsens from 0.300 to 0.319 mm. Projection is post-processing, not proof of agronomic accuracy.
Because the exact synthetic target lies inside the projected feasible set, projection cannot increase total squared one-step error. That reduction is partly guaranteed by geometry, not evidence of learned physics. Individual-output and free-running errors need not improve. In fact, projection worsens rollout storage error for seed 202 in 2022 and 2023; the raw model reaches −29.62 mm storage.
All tested projected outputs obey the selected bounds; maximum one-step balance residual is below 5×10⁻¹⁴ mm. This checks arithmetic constraints, not the adequacy of the underlying crop physics.
Evidence and limits
- Training 2001–2015; validation 2016–2018; retrospective evaluation 2019–2024. Evaluation years were examined in earlier analyses, so this is exploratory.
- All crop states and targets are synthetic. No observed soil, irrigation, yield or sales targets are available.
- Satellite imagery remains geographic context, not an input feature. Vegetation data requires matching, quality controls and field outcomes before a crop-value claim.
- The exact bucket remains the scenario engine. No speed advantage or reason to replace it is established.
Supporting tables: validation.csv, evaluation.csv, rollout.csv and rollout_daily.csv. Root mean square error (RMSE) is a typical numerical error magnitude. The displayed averages are across three seeds, not confidence intervals.
Terms explained
Select a term to read its meaning in this study.
Reanalysis
An estimate of past weather made by combining observations with a weather model. It is not a measurement at every farm.
Monte Carlo simulation
A comparison repeated with many sampled inputs to show how results can change when conditions are uncertain.
Baseline
The reference plan or model that another option is compared with.
Water balance
An accounting of the water entering, leaving and remaining in the model.
Neural surrogate
A trained model that tries to imitate another calculation. Here it is compared with a simple water-storage simulator, not measured crop yields.
Physics-informed penalty
An extra training cost for breaking a stated physical relationship. A smaller penalty does not guarantee a more accurate answer.
Root mean square error
A measure of the typical size of prediction differences that gives extra weight to large differences. It must be read with its units and comparison target.
Projection
A mathematical adjustment that moves a prediction into a set of allowed values. Satisfying those rules does not prove that a crop response is correct.
Validation data
Data used to choose model settings. A separate evaluation set is needed to check the chosen model.
Retrospective check
A comparison using past data. It does not by itself show what would have been known when a real decision was made.