Real-Time Climate Risk Assessment for Supply Chain Resilience: A Data-Driven Nowcasting Framework for Colombian Agriculture

Agricultural supply chains in Colombia are highly vulnerable to climate variability, yet many existing early warning systems rely on satellite imagery or complex remote sensing infrastructure that is often inaccessible or computationally prohibitive. This paper addresses this gap by proposing a methodological framework that integrates short-term climate nowcasting with supply chain risk modeling. The approach uses ground-based meteorological station data and official agricultural statistics to generate actionable risk signals for time-sensitive crops like coffee, rice, and flowers.

The framework operates through three primary stages: data preprocessing, climate nowcasting, and risk mapping. First, historical meteorological data (e.g., temperature, precipitation, humidity) and agricultural yield statistics are cleaned and normalized. Second, a Long Short-Term Memory (LSTM) neural network is trained to predict precipitation and temperature conditions at 6, 12, 24, and 48-hour horizons. Finally, these forecasts are compared against crop-specific, threshold-based risk categories to provide supply chain managers with clear, actionable signals for decision-making regarding sourcing, inventory, and transport.

The prototype implementation, tested using synthetic data calibrated on Colombian climate patterns, demonstrates that the LSTM model maintains stable error metrics (MAE 0.58–0.60 mm) across all forecast horizons. Notably, the model's ability to detect extreme weather events—measured by F1-score—actually improves at longer lead times (reaching 0.66 at 48 hours), suggesting the architecture effectively captures underlying seasonal patterns. The study confirms that these nowcasts can be successfully translated into categorical risk indicators, providing a viable, low-infrastructure alternative for climate-adaptive supply chain management in developing regions.