CO₂ Emissions in
Agro-food Systems

Problem Statement

The agro-food sector is a major contributor to greenhouse gas emissions across Latin America, yet emission patterns remain poorly understood at the country level. This project analyzes CO₂ emissions from agro-industrial activities in IDB member countries using an annual, country-level dataset covering 1990–2020.

The central question: Can historical patterns in agro-food activity variables reliably explain and project total CO₂ emissions toward 2030? The goal is an interpretable, data-driven framework that supports emissions monitoring and scenario-based exploration — without assuming causal relationships or prescribing policy.

Analytical Pipeline

Data Overview

The dataset contains 13 columns covering individual emission sources — food processing, packaging, transport, retail, household consumption, rice cultivation, on-farm energy use, and fertilizer/pesticide manufacturing — plus the aggregate total_emission target. All numeric; no duplicates.

Emission Trends & Structure

Regional Trajectory (1990–2020)

Total agro-food emissions grew steadily from ~1.2M kt in 1990 to ~1.9M kt in 2020 — a 58% increase over 30 years. Food Household Consumption and Food Transport are the largest and fastest-growing components, reflecting regional urbanization and food system expansion. A structural acceleration is visible around 2010–2012.

Emissions by Country

Emissions are highly geographically concentrated. Brazil alone accounts for ~40% of the regional total — a single country whose trajectory drives global metrics. Argentina and Mexico form a distant second tier (~7–8% each). This heterogeneity is fundamental to interpreting model performance.

Features Constructed

Three types of features are constructed from the raw variables:

Aggregated sector variables: food_chain_total (processing + packaging + transport + household + retail) and on_farm_total (energy use + electricity). These reduce dimensionality while preserving the main structural signals and directly address the multicollinearity observed in the correlation matrix.

Lagged emissions (total_emission_lag1): captures temporal autocorrelation within each country. Computed per country via groupby("Area").shift(1) — panel-safe, prevents cross-country leakage.

Year: encodes the overall time trend common across countries.

# Panel-safe feature construction
df = df.sort_values(["Area", "Year"]).reset_index(drop=True)

df["food_chain_total"] = (
    df["Food Processing"] + df["Food Packaging"] +
    df["Food Transport"]  + df["Food Household Consumption"] +
    df["Food Retail"]
)
df["on_farm_total"] = df["On-farm energy use"] + df["On-farm Electricity Use"]

# Lag computed within each country — avoids cross-country leakage
df["total_emission_lag1"] = df.groupby("Area")["total_emission"].shift(1)

features = ["Rice Cultivation", "food_chain_total", "on_farm_total",
            "total_emission_lag1", "Year"]

Three Complementary Models

A time-based split (train ≤ 2015, test 2016–2020) is used throughout to preserve temporal ordering and prevent leakage. This yields 575 training rows and 115 test rows. The cutoff is set at 2015 — not at the observed structural break (~2010) — to retain a minimum of 5 full test years sufficient for stable metric estimation.

Feature Importance (Tree-based models)

Both RandomForest and GradientBoosting agree on the ranking of features, providing a consistent signal about which variables drive emissions:

How Much Do Activity Variables Actually Contribute?

The lagged emission variable (total_emission_lag1) is a very strong predictor because it encodes the same quantity being predicted, shifted one period. A high R² may therefore reflect temporal autocorrelation rather than genuinely predictive relationships between agro-food activities and emissions.

To assess the independent contribution of the activity variables, all three models are trained both with and without the lag feature. This ablation makes the analytical contribution of the project honest and defensible.

Overall Model Comparison (with lag)

ElasticNet is the best model on all three metrics. GradientBoosting's RMSE is 3× higher, consistent with tree-based models reverting toward the training mean on out-of-sample periods — particularly for countries whose 2016–2020 trajectory diverges from the 2015 trend. The per-country validation confirms this: the countries where GBR error is highest are the same small-economy cases with greatest distributional shift. ElasticNet's interpretability also makes it the right choice for scenario projections.

Conditional Forecast to 2030

This section presents a hypothetical scenario. Driver trajectories are derived directly from observed CAGR over 2011–2020 (code below). Results are conditional projections, not deterministic forecasts — changing any growth rate substantially alters the outcome.

# Reproducible derivation of scenario growth rates from data
base_year, end_yr = 2011, 2020
n = end_yr - base_year  # 9 years

for col in ["Rice Cultivation", "food_chain_total", "on_farm_total"]:
    v0 = df.loc[df["Year"] == base_year, col].sum()
    v1 = df.loc[df["Year"] == end_yr,   col].sum()
    cagr = (v1 / v0) ** (1 / n) - 1
    print(f"{col:25s}  CAGR = {cagr:+.1%}")

The projection is iterative: each year's predicted emission feeds the next year as total_emission_lag1. Driver values are updated by compounding the CAGR from the last observed per-country value. The model used for projection (enet_final) is retrained on all 1991–2020 data; the model used for reported metrics (enet_eval, trained on ≤2015 only) is kept separate.

Per-Country Model Performance

Global metrics hide important heterogeneity. Brazil alone generates ~40% of regional emissions — a model that fits Brazil perfectly can achieve high global R² while performing poorly on smaller countries. The table below computes Relative MAE (MAE as % of country mean) for each country on the test period 2016–2020.

Summary & Limitations

Main Insights

Agro-food CO₂ emissions across IDB countries grew from ~1.2M kt (1990) to ~1.9M kt (2020), driven primarily by the food supply chain and on-farm energy use. Emissions are highly geographically concentrated — Brazil alone accounts for ~40% of the regional total. The lagged emission variable is the single strongest predictor (~60–75% of tree-based importance), confirming strong temporal inertia. The ablation study reveals that agro-food activity variables contribute an additional ~15 percentage points of R² — a meaningful and independently important finding.

Model Performance

ElasticNet outperforms tree-based models across all metrics on the temporal test period (2016–2020): R²=0.993, MAE≈8,064 kt, RMSE≈14,119 kt. Per-country validation shows strong performance for major emitters (R²>0.95) and weaker performance for small island states (R²=0.74–0.83), where idiosyncratic dynamics dominate.

Limitations