Shi et al. (2025) Ai-driven spatiotemporal modelling of agricultural biomass residues in Northern Italy: A data-fusion and explainable-AI framework

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2025
• Date: 2025-11-21
• Authors: Zhan Shi, Francesco Marinello, Andrea Pezzuolo
• DOI: 10.1016/j.compag.2025.111246

Research Groups

• Department of Land, Environment, Agriculture and Forestry, University of Padova, Legnaro (PD) 35020, Italy
• Department of Agronomy, Food, Natural Resources, Animals and Environment, University of Padova, Legnaro (PD) 35020, Italy

Short Summary

This study presents an AI-based framework integrating Feedforward Neural Networks and XGBoost with remote sensing and explainable AI to assess and forecast agricultural biomass residue potential across 47 provinces in Northern Italy from 2002 to 2031. It finds that maize and wheat dominate residue supply, contributing over 100 million metric tonnes of dry biomass annually, but forecasts a decline in productivity after 2026 due to rising temperatures, increased water deficits, and land-use change.

Objective

• How can machine learning models, combined with spatial datasets, enhance provincial-level predictions of residue-based biomass productivity and bioenergy potential?
• Which climatic, agronomic, and economic factors have historically shaped bioenergy production from crop residues in Northern Italy, and how do these key drivers vary across provinces?
• How will bioenergy potential from crop residues evolve under forecasted climate and land-use scenarios, and what does this imply for long-term energy planning and sustainable transitions?
• How can these AI-driven insights guide adaptive resource management and policy strategies, given the spatial diversity, to maintain resilient bioenergy systems?

Study Configuration

• Spatial Scale: 47 provinces in Northern Italy (NUTS 3 Level). Crop mapping data at approximately 10 meters and 5 kilometers resolution. Net Primary Productivity (NPP) data at 500 meters resolution.
• Temporal Scale: Historical data from 2002 to 2022. Forecasts for 2023 to 2031. Data collected at monthly and yearly frequencies.

Methodology and Data

• Models used:
  • Machine Learning: Feedforward Neural Networks (FNN), XGBoost, Decision Tree (DT), Random Forest (RF), Support Vector Regression (SVR).
  • Baseline: Linear Regression (LR).
  • Explainable AI: SHAP (SHapley Additive exPlana-tions), LIME (Local Interpretable Model-agnostic Explanations).
  • Forecasting: Holt-Winters method.
• Data sources:
  • Satellite: MODIS (Net Primary Productivity (NPP), Gross Primary Production (GPP)), EUCROPMAP, CROPGRIDS, EUcropmap (crop mapping data).
  • Observation/Reanalysis: TerraClimate Monthly dataset (climate variables including temperature (K), precipitation accumulation (mm), vapor pressure (kPa), wind speed at 10 m (m/s), soil moisture, actual/reference evapotranspiration (mm), climate water deficit (mm), snow water equivalent (mm), downward shortwave radiation (W/m²), Palmer Drought Severity Index).
  • Statistical: ISTAT (crop planting areas (km²), crop production (Mg), plant protection products (Mg), fertilizer (Mg), agricultural producer price (EUR), renewable energy electricity generation (GWh), provincial GDP, environmental protection data, population, territorial area (km²)).
  • Emissions: EDGAR database (GHG emissions: CH₄, CO₂, CO₂-bio, N₂O, F-gases in metric tonnes of CO₂ equivalent (Mg CO₂eq)).

Main Results

• The FNN and XGBoost models demonstrated strong predictive performance for NPP (R² > 0.89, relative Mean Absolute Error (rMAE) between 8% and 18%).
• Maize and wheat consistently provide the highest dry residue biomass, contributing over 100 million metric tonnes (Mg) annually.
• The total bioenergy potential from crop residues in Northern Italy is estimated to be approximately 960–1,100 billion MJ annually, supporting 95–100 billion kWh of electricity.
• Energy density from crop residues peaks at 30 MJ/m² in the Po Valley.
• Explainable AI (SHAP and LIME) identified climate water deficit, soil moisture, evapotranspiration, crop area, and fertilizer input as key drivers of NPP. Maize NPP is predominantly climate-water controlled, while wheat NPP shows stronger sensitivity to management and water availability.
• Forecasts beyond 2026 project a decline in biomass productivity and bioenergy potential (from 10.2 MJ/m² to 8.4 MJ/m² by 2031) due to rising temperatures, increased water deficits, and land-use changes.
• High-yield provinces (e.g., Mantova, Alessandria, Cremona, Lodi, Milano, Piacenza) are projected to experience significant reductions in accumulated NPP, up to 49%, while moderate and low-productivity areas show more stability or persistent constraints.
• The Po Valley remains the dominant bioenergy production zone, but its potential is projected to decline after 2026.

Contributions

• Developed a novel data-fusion and explainable-AI framework that integrates advanced machine learning models (FNN, XGBoost) with remote sensing and interpretability techniques (SHAP, LIME) for high-resolution, long-term assessment and forecasting of agricultural biomass residues.
• Provided the first provincial-level, long-term (2002-2031) assessment of agricultural biomass residue availability and bioenergy potential in Northern Italy under evolving climate and land-use conditions, addressing a critical gap in fine spatiotemporal scale analysis.
• Quantified the influence of key climatic, agronomic, and land-use factors on biomass dynamics using explainable AI, enhancing model transparency and trust for data-driven decision-making.
• Offered forward-looking insights into future bioenergy supply under climate change scenarios, highlighting regional vulnerabilities and providing a basis for climate-resilient agriculture and regional resource planning.
• Advanced smart agriculture by integrating machine learning, remote sensing, and model interpretability to improve biomass monitoring and support long-term sustainability strategies.

Funding

• European Union (Grant Agreement number: 101118296)

Citation

@article{Shi2025Aidriven,
  author = {Shi, Zhan and Marinello, Francesco and Pezzuolo, Andrea},
  title = {Ai-driven spatiotemporal modelling of agricultural biomass residues in Northern Italy: A data-fusion and explainable-AI framework},
  journal = {Computers and Electronics in Agriculture},
  year = {2025},
  doi = {10.1016/j.compag.2025.111246},
  url = {https://doi.org/10.1016/j.compag.2025.111246}
}