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

This paper presents a methodological framework for real-time climate risk assessment using data-driven nowcasting techniques to enhance supply chain resilience in Colombian agricultural contexts. Climate variability in Colombia, characterized by irregular rainfall, temperature fluctuations, and recurrent extreme events, has a direct impact on agricultural production and logistics, particularly for time sensitive crops. The proposed approach integrates short term climate forecasting based on historical meteorological observations with supply chain risk modeling to establish a conceptual early warning system architecture. A prototype implementation developed in a controlled computational environment demonstrates the feasibility of the framework using historical meteorological and agricultural time series derived from official statistics and reanalysis products, without reliance on satellite imagery or computer vision components. The methodology addresses the integration of climate nowcasting with supply chain decision making through explicit risk mapping, threshold-based categorization, and stakeholder-oriented risk signals. Results from synthetic and historical data experiments indicate that short term precipitation nowcasts can be translated into actionable risk indicators for agricultural supply chains, supporting anticipatory decisions related to inventory, sourcing, and transport.

The paper proposes a methodological framework for real-time climate risk assessment in supply chains through the integration of nowcasting techniques, climate-agriculture relationships, and supply chain risk modeling. The approach leverages historical ground-based meteorological observations and official agricultural statistics to establish baseline climate–production relationships, which are then translated into threshold-based risk categories and early warning signals. A prototype implementation is developed and validated in a controlled computational environment using synthetic scenarios calibrated on documented Colombian climate variability and crop production patterns, ensuring logical consistency and technical feasibility before operational deployment. The framework is designed to operate with datasets readily available from Colombian national institutions and internationally accessible reanalysis products, establishing a foundation for future field deployment and refinement through collaboration with IDEAM, the Ministry of Agriculture, and supply chain stakeholders in vulnerable regions.

The contribution of this work relies on the development of a structured methodology for short-term climate nowcasting with supply chain risk assessment using time series data, explicit threshold-based risk mapping and a proof of prototype that demonstrates technical and operational viability of the framework. While field deployment and stakeholder validation remain future work, the theoretical and methodological foundation presented here addresses critical gaps in climate-adaptive supply chain management for developing country contexts where advanced remote sensing infrastructure is limited.

The proposed methodology follows three primary concepts: climate nowcasting through deep learning applied to ground-based meteorological time series, climate-agriculture impact mapping based on historical relationships and supply chain risk translation into operationally interpretable signals. The framework is designed to operate within a computational environment, with potential for deployment to production systems serving multiple stakeholders. The framework is designed to operate with data sources typically available in Colombian agricultural contexts, including meteorological observations from IDEAM and official agricultural statistics from the Ministry of Agriculture. The workflow processes historical time series to establish baseline relationships between climate variables and production outcomes, trains predictive models on ground-based observations, and generates forward looking risk signals aligned with supply chain planning horizons measured in hours to days.

The framework integrates meteorological and agricultural data streams into a unified analytical dataset with daily precipitation, temperature (minimum, maximum, mean), relative humidity, solar radiation and vapor pressure deficit sourced from IDEAM meteorological stations and aggregated to agricultural zones. Quality control procedures include outlier detection using interquartile range methods (IQR), missing value imputation via temporal interpolation, and consistency checks against neighboring stations to ensure data integrity. Regional yield statistics for key crops (coffee, rice, flowers) are obtained from official agricultural ministry databases and AGRONET, the national agricultural statistics portal, harmonized to monthly or seasonal aggregation to align with the temporal resolution of meaningful crop response cycles. Standardized Precipitation Index (SPI) and temperature anomalies are computed from raw meteorological data to capture drought and heat stress conditions, enhancing the model's ability to represent slow-onset climate variations. All data streams are normalized to zero mean and unit variance prior to model training, enabling stable convergence of neural network architectures and preventing feature dominance due to differing scales.

The nowcasting model predicts short term climate (6-48 hour) precipitation and temperature conditions using recent observational history, formulated as a supervised time series prediction problem where a sliding window of past meteorological observations (typically 24-48 hours) is used to predict conditions 6, 12, 24, and 48 hours ahead at the regional level. A Long Short-Term Memory (LSTM) recurrent neural network is selected as the primary architecture, building on demonstrated effectiveness in precipitation nowcasting. The LSTM captures long-term dependencies in time series data through its memory cell mechanism, enabling the model to learn which historical states are relevant for future predictions. While alternative architectures such as CNN-LSTM hybrids are considered for spatial-temporal feature extraction when multi-station or gridded data are available, a single-region LSTM is retained for the initial prototype due to computational simplicity, interpretability, and suitability for station-based input data without satellite imagery.

The architecture consists of time-lagged meteorological features (temperature, rainfall, humidity, pressure) as input, two stacked LSTM layers with 64 and 32 hidden units respectively with dropout regularization (p=0.2) to prevent overfitting, and a dense output layer producing scalar predictions for each forecast horizon, optimized using Adam (learning rate = 0.001) with Mean Squared Error loss. A time aware cross validation scheme is employed to preserve temporal order and prevent information leakage, with historical data partitioned into training (2017-2021), validation (2022), and test (2023-2024) sets. This partition reflects the temporal progression of IDEAM and AGRONET records and ensures that validation and test periods do not influence model training. Models are trained for up to 100 epochs with early stopping based on validation loss, minimizing the risk of overfitting to local patterns.

To translate nowcasting outputs into supply chain relevant signals, empirical relationships between climate variables and agricultural outcomes are established by computing Pearson and Spearman correlations between lagged meteorological variables and yield anomalies for each region and crop, identifying key climate drivers. Risk thresholds are defined using historical quantiles of meteorological variables and their documented impacts on crop production. For precipitation-driven risks, thresholds categorize conditions as Low, Moderate, or High based on the intensity and duration of anomalies derived from IDEAM historical records. For drought-affected crops, cumulative water deficits during sensitive growth stages define impact severity, informed by agronomic literature and Ministry of Agriculture guidance on crop phenology. Thresholds are crop and region specific, informed by agronomic literature, observational data, and consultation with documented crop requirements in Colombian agroecological zones. For each time step and region, nowcasting outputs are compared against defined thresholds to generate categorical risk levels that explicitly incorporate lead time (hours to days ahead), enabling supply chain managers to adjust inventory, sourcing, or transportation planning before disruption materialization.

The entire framework is implemented in a Google Colab environment using Python libraries (Pandas for data manipulation, Keras/TensorFlow for neural networks, Scikit-learn for preprocessing). This controlled computational configuration allows for iterative model development, validation against historical IDEAM and AGRONET data, and systematic evaluation of framework components without requiring field deployment infrastructure. The prototype processes synthetic historical scenarios calibrated on documented Colombian climate patterns and official agricultural production statistics, validating the framework's logical consistency and decision relevance before field-level deployment. The use of synthetic calibrated data at this stage is methodologically justified, as it allows controlled testing of framework architecture while maintaining consistency with real Colombian climate and agricultural patterns documented in IDEAM and Ministry of Agriculture records.

Framework performance is assessed through two complementary indexes: nowcasting accuracy using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and F1-score for extreme event detection; and risk signal utility through qualitative assessment of whether generated risk categories align with observed patterns in agricultural outcomes and whether lead times are sufficient for supply chain adaptation actions.

The development utilized synthetic meteorological data calibrated on documented Colombian climate patterns from IDEAM historical records (2017–2024) and agricultural production data aligned with official statistics from AGRONET and the Ministry of Agriculture. Three representative Colombian agricultural regions were selected for framework validation: the Andean Coffee Belt (Huila, Cauca, Nariño), the Rice-Producing Zone (Tolima, Córdoba), and the Inter-Andean Region (Cundinamarca, Boyacá). These regions account for approximately 65% of Colombia's agricultural GDP and represent the primary crops most sensitive to climate variations: coffee, rice, and maize. The synthetic dataset comprised 2,922 daily meteorological observations (January 2017 – December 2024) and 96 monthly agricultural records spanning the same period, calibrated on documented Colombian climate seasonality and crop production patterns.

The LSTM model was trained on the 2017-2021 period (1461 observations), validated on 2022 (365 observations), and tested on 2023-2024 (731 observations). The model demonstrates consistent performance across forecast horizons, with MAE and RMSE remaining stable (0.58–0.60 mm and 0.73–0.75 mm respectively) due to the normalization applied during preprocessing. F1-scores for extreme event detection improve at longer forecast horizons (6h: 0.51 → 48h: 0.66), indicating that the model's discriminatory power for identifying high-risk precipitation events strengthens at extended lead times. This counterintuitive result suggests that the model captures underlying temporal patterns in extreme event occurrence more reliably over extended windows, possibly reflecting dominant seasonal signals in Colombian rainfall patterns. The 12-hour horizon achieves F1-score of 0.53, providing moderate discriminatory power for operational risk categorization. The consistent low error metrics across all horizons indicate that the synthetic data calibration successfully captured normalized meteorological variability, and that the LSTM architecture generalizes effectively to unseen validation periods.

Empirical correlations between lagged meteorological variables and regional yield anomalies confirmed documented climate sensitivities. The analysis revealed heterogeneous climate sensitivities across crop types, reflecting documented agronomic responses to meteorological forcing. Mean temperature showed moderate positive correlation with coffee yield (r = 0.38), suggesting that within the historical temperature range observed in the Andean Coffee Belt (15–19°C), warmer conditions within documented optimal ranges (16–21°C) are associated with improved productivity. This moderate strength reflects the non-linear nature of temperature-yield relationships, where benefits occur only within narrow optimal windows temperatures exceeding crop-specific thresholds produce yield reductions. Monthly precipitation showed weak correlation with rice production (r = 0.09), indicating that the synthetic dataset calibration may not fully capture the acute sensitivity of rice to water availability during specific critical growing stages. Rice yield is highly responsive to irrigation management and intra-seasonal water stress timing rather than monthly aggregate precipitation; the weak aggregate correlation reflects this phenological specificity. Flower yields demonstrated strong positive correlation with mean temperature (r = 0.66), consistent with documented temperature sensitivity of ornamental crop production in Sabana de Bogotá regions, where temperature fluctuations directly affect flowering phenology, stem elongation rates, and marketable yield.

Risk categories were defined using historical quantiles of meteorological variables derived from the synthetic dataset, calibrated on documented Colombian climate patterns and agricultural vulnerability literature. Thresholds were established at the 33rd and 67th percentiles of the empirical distribution of meteorological anomalies to create balanced risk categories. For the Andean Coffee Belt, the 48-hour cumulative rainfall deficit thresholds (Low: 44.4 mm) reflect the characteristic intensity and frequency of precipitation events in Colombian highland regions. The moderate threshold of 29.6–44.4 mm deficit corresponds to incipient water stress conditions that trigger increased monitoring and inventory pre-positioning, while the high threshold (> 44.4 mm deficit over 48 hours) represents severe drought conditions requiring immediate contingency activation. Temperature anomaly thresholds (Low: 2.0°C) are narrower, reflecting the documented sensitivity of coffee productivity to temperature deviations within the optimal growing range. The tight temperature thresholds align with the moderate correlation (r = 0.38) observed between mean temperature and coffee yield, indicating that even small departures from optimal conditions warrant supply chain precautions.

The three-stage translation process operated consistently across prototype iterations. Nowcasting outputs successfully linked to agricultural outcomes through defined thresholds, demonstrating logical coherence of the framework. Qualitative assessment confirmed that generated risk signals could feasibly inform supply chain planning decisions regarding inventory pre-positioning, sourcing diversification, and transportation rerouting, though stakeholder feedback validation remains future work.

The proposed framework introduces several methodological advances relevant to developing-country agricultural supply chain contexts. First, it demonstrates operationalization of nowcasting using only ground-based meteorological stations and official agricultural statistics, circumventing remote sensing infrastructure requirements that limit applicability in resource-constrained settings. IDEAM's relatively dense station network across agricultural regions and AGRONET's standardized production records make this approach immediately implementable in Colombian institutional contexts. Second, the quantile-based threshold derivation (33rd and 67th percentiles) provides a principled, reproducible approach to risk categorization aligned with historical climatology. For the Andean Coffee Belt, this yielded parsimonious thresholds (rainfall def…

Improvements to AI Systems:

1. Climate-Nowcasting AI with LSTM-Based Temporal Prediction: Integrate the paper’s LSTM architecture (two stacked layers with 64 and 32 hidden units, dropout 0.2, Adam optimizer) to build AI systems that predict precipitation and temperature at 6–48 hour lead times using only ground-based meteorological time series, eliminating reliance on satellite imagery. The improved system can generate short-term climate forecasts with stable MAE (0.58–0.60 mm) and RMSE (0.73–0.75 mm), and detect extreme events with F1-scores up to 0.66 at 48 hours.

2. Threshold-Based Risk Categorization Engine: Implement the quantile-based threshold derivation (33rd and 67th percentiles) to create AI systems that automatically classify climate conditions into Low/Moderate/High risk categories. The improved system can produce parsimonious, reproducible thresholds (e.g., rainfall deficit 29.6–44.4 mm, temperature anomaly 1.9–2.0°C for coffee) and generate actionable risk signals with a 38-hour mean lead time, enabling anticipatory supply chain decisions.

3. Crop-Specific Climate-Yield Correlation Modeling: Use the empirical correlation analysis (Pearson/Spearman) to build AI systems that learn heterogeneous climate sensitivities across crops—e.g., strong temperature coupling for flowers (r=0.66), moderate for coffee (r=0.38), weak aggregate precipitation for rice (r=0.09). The improved system can differentiate risk assessment by crop type and region, avoiding one-size-fits-all predictions and capturing phenological specificity.

4. Time-Aware Cross-Validation and Early Stopping: Adopt the paper’s temporal partitioning scheme (train 2017–2021, validate 2022, test 2023–2024) and early stopping (up to 100 epochs) to build AI systems that prevent information leakage and overfitting. The improved system can generalize to unseen periods while preserving temporal order, ensuring robust performance in real-world deployment.

5. Data Quality Control and Preprocessing Pipeline: Integrate IQR-based outlier detection, temporal interpolation for missing values, and consistency checks against neighboring stations to create AI systems that handle noisy, incomplete ground-based data. The improved system can normalize features to zero mean and unit variance, compute SPI and temperature anomalies, and harmonize multi-source data (IDEAM, AGRONET) into unified analytical datasets.

6. Risk Signal Translation for Supply Chain Decisions: Implement the three-stage translation process (nowcasting → climate-agriculture mapping → risk signals) to build AI systems that convert meteorological outputs into operationally interpretable signals. The improved system can inform inventory pre-positioning, sourcing diversification, and transportation rerouting, with a three-level categorical structure that reduces cognitive burden for stakeholders.

7. Modular and Transferable Framework Architecture: Use the paper’s modular design (crop-specific and region-specific customization) to create AI systems that adapt to diverse agricultural contexts beyond Colombia. The improved system can be reconfigured for different crops, regions, and data availability, making it transferable to other developing countries with limited remote sensing infrastructure.

8. Synthetic Data Calibration for Prototype Validation: Apply the paper’s method of calibrating synthetic data on documented climate patterns (e.g., 2,922 daily observations, 96 monthly records) to build AI systems that validate framework logic before field deployment. The improved system can test architecture consistency and decision relevance without requiring operational infrastructure, reducing development risk.

9. Extreme Event Detection with Extended Lead Times: Leverage the finding that F1-scores improve at longer horizons (0.51 at 6h → 0.66 at 48h) to build AI systems that prioritize high-risk event prediction over extended windows. The improved system can provide more reliable early warnings for severe drought or flood conditions, enabling longer preparation windows for supply chain managers.

10. Re-Calibration and Continuous Learning Mechanism: Incorporate the paper’s recommendation for periodic threshold re-calibration as new observations accumulate to build AI systems that adapt to climate change and shifting seasonality. The improved system can update risk thresholds and model weights dynamically, maintaining accuracy under non-stationary environmental conditions.

This paper presents a methodological framework for real-time climate risk assessment using data-driven nowcasting techniques to enhance supply chain resilience in Colombian agricultural contexts. Climate variability in Colombia, characterized by irregular rainfall, temperature fluctuations, and recurrent extreme events, has a direct impact on agricultural production and logistics, particularly for time sensitive crops. The proposed approach integrates short term climate forecasting based on historical meteorological observations with supply chain risk modeling to establish a conceptual early warning system architecture. A prototype implementation developed in a controlled computational environment demonstrates the feasibility of the framework using historical meteorological and agricultural time series derived from official statistics and reanalysis products, without reliance on satellite imagery or computer vision components. The methodology addresses the integration of climate nowcasting with supply chain decision making through explicit risk mapping, threshold-based categorization, and stakeholder-oriented risk signals. Results from synthetic and historical data experiments indicate that short term precipitation nowcasts can be translated into actionable risk indicators for agricultural supply chains, supporting anticipatory decisions related to inventory, sourcing, and transport.