From Rainfall Predictions to Crop Simulations:
Model Validation in Climate-Data Pipelines

Introduction

Why Simulate Crop Growth?

• Crop models as tools for climate resilience

• Perspectives:

  • Farmer’s Perspective: managing volatility, extreme weather, planting windows

  • Business Perspective: forecasting for risk management, insurance, logistics

  • Food Security Perspective: early-warning systems and vulnerability mapping

  • Policy & Research Perspective: evidence-based adaptation and investment planning

• Integration of rainfall predictions makes crop models dynamic resilience tools

• Reliability of process-based crop models: mechanistic, data-driven, integrative

• Features of reliable models:

  • Mechanistic (physics, chemistry, biology)

  • Integration across the soil–plant–atmosphere continuum

  • Scalability and uncertainty quantification

The Climate Data Pipeline

• Overview: From data → model → simulation → decision → impact

• Data (Foundation):

  • Inputs: weather (temperature, solar radiation, precipitation), soil, management

  • Challenges: noise, inconsistency, coarse resolution

• Model (Processing Engine):

  • Process-based equations for photosynthesis, transpiration, growth, yield

  • Requires calibration and validation

• Simulation: thousands of runs → probabilistic outcomes

• Decision: translating results into farmer, business, or policy action

• Impact: economic, environmental, and social implications

• Validation as both a scientific and ethical responsibility

Why This Matters

• Example: Typhoon rainfall forecasting and its role in crop-management decisions

• Reference: Zhou et al. (2023) on global concurrent climate extremes

Climate Data Realities — Imbalance & Noise

• Imbalance: many “normal” days, few “extreme” events

• Noise: sensor error, missing values, measurement uncertainty

• Case Example: Taal Lake Fishkill (Pabico et al. 2015)

  • Many “no-fishkill” vs. rare “fishkill” days

  • Missing and irregular data issues

Imbalanced Data in Modeling

• Standard accuracy misleads under imbalanced classes

• Alternative metrics: precision, recall, F1, PR-curve (Miftahushudur et al. 2025)

• Resampling Techniques:

  • Undersampling: random, edited nearest neighbor, cleaning rule, Tomek links

  • Oversampling: random, SMOTE, ADASYN, Borderline SMOTE, SL-SMOTE, K-means SMOTE, SVM SMOTE

SMOTE — Synthetic Minority Oversampling Technique

• Reference: Chawla et al. (2002)

• Five Steps:

  • Identify minority class

  • Select sample + nearest neighbors

  • Randomly choose neighbor

  • Create synthetic sample

  • Repeat until classes balanced

Noise & Data Quality Issues

• Rainfall measurement errors propagate downstream (Heinemann et al. 2002)

• Sources: gauge design, aerodynamic effects, wetting losses

• Effects on simulated soil moisture, evapotranspiration, yield

• Model calibration cannot compensate for poor rainfall data

Interpreting Heinemann et al. (2002) Through a DS Lens

• Rainfall time series are inherently imbalanced

• Rare events (heavy rainfall) drive model sensitivity

• Measurement errors on rare days disproportionately affect outcomes

• Lesson: validation must emphasize rare, high-impact events

“Event Imbalance” as Hidden Bias

• Imbalance between normal and extreme rainfall affects yield predictions

• Conceptual parallel with minority-class misclassification in ML

• Example mapping between data bias and crop-model implications

Reflections

• Imbalance appears as temporal or spatial unevenness

• Ethical duty to represent rare, high-impact conditions

• When extremes are underrepresented, both AI and crop models learn an incomplete reality

Supporting Study: Thaler et al. (2018)

• Spatial variability of precipitation affects crop simulations

• High-resolution local data yields better outputs

• Interpolation smooths extremes, underestimating water stress

• Satellite rainfall misrepresents local variability

• Importance of spatial + temporal resolution in validation

Bonus — Encoding Quantitative & Periodic Variables

• Encoding continuous and cyclic variables (Pabico 1996)

• Interpolation representation for continuous ranges (e.g., growth −20% → 40%)

• Example: proportional encoding between adjacent nodes

• Periodic variables (time, date) require circular encoding

• Wrap-around representations for cyclic continuity

Closing

• Invitation to questions and open discussion: “Audience's puzzling questions and speaker's (hopefully) logical answers.”