🤖 machine learning

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

This paper proposes and validates a data-driven nowcasting framework that integrates short-term climate forecasts with supply chain risk modeling to enable real-time, actionable risk assessment and anticipatory decision-making for Colombian agricultural supply chains without relying on satellite imagery.

Original authors: Hernan J. Silva-Sosa

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hernan J. Silva-Sosa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive ship, but instead of sailing on the ocean, you are navigating a river of fresh produce. Your cargo holds are filled with coffee beans, rice, and flowers, all waiting to be shipped to markets around the world. The problem? The weather is a mischievous trickster. In places like Colombia, rain can suddenly turn into a flood, or the sun can bake the fields dry, all within a few hours. If you don't know the weather is changing until you see the storm clouds, it's often too late to save your cargo.

To solve this, scientists use something called "nowcasting." Think of it as a super-accurate crystal ball that doesn't look months into the future, but rather predicts the next few hours or days with high precision. It's like checking the radar on your phone to see if you need an umbrella in the next hour, rather than a weather report for next month. When you combine this with "supply chain resilience," you get a system that helps businesses prepare for the storm before it hits. Instead of just reacting when the rain starts, they can move their inventory, change their routes, or call in extra help while the sky is still clear. This paper explores how to build such a system for farmers and food companies in Colombia, using only the data they already have on the ground, without needing expensive satellites or fancy cameras.

The Paper's Mission: A Weather Oracle for Food Trucks

This paper introduces a clever new framework designed to act as a real-time "weather oracle" for Colombia's agricultural supply chains. The author, led by Hernan J. Silva-Sosa, wanted to see if they could build a system that takes short-term weather predictions and instantly translates them into actionable advice for businesses. The goal is to help companies decide whether to stock up on goods, reroute their trucks, or switch suppliers before a climate disaster actually happens.

The paper explicitly rules out the idea that you need high-tech satellite imagery or complex computer vision (like AI that "sees" clouds from space) to make these predictions. Instead, the author argues that you can get the job done using only data from ground-based weather stations and official government statistics. They built a prototype to prove this concept works, but it's important to note that this was a simulation run in a controlled computer environment, not a live system tested on real farms yet.

How the Machine Thinks

The core of their invention is a digital brain powered by a type of artificial intelligence called an LSTM (Long Short-Term Memory) network. You can think of this AI as a very attentive student who has studied the history of weather patterns for years. It looks at the last 24 to 48 hours of data—temperature, humidity, and rain—and tries to guess what the weather will be like in 6, 12, 24, and even 48 hours.

In their tests, this digital student performed quite consistently. When predicting rain, the error was very small, hovering between 0.58 mm and 0.60 mm (Mean Absolute Error) and 0.73 mm to 0.75 mm (Root Mean Square Error). Perhaps most surprisingly, the AI got better at spotting extreme weather events the further out it looked. Its ability to correctly identify a "high-risk" storm improved from an accuracy score (F1-score) of 0.51 at 6 hours ahead to 0.66 at 48 hours ahead. This suggests that for this specific region, the patterns of extreme weather are actually easier to spot when looking at a slightly longer window.

Turning Rain into Decisions

Knowing it's going to rain is useful, but knowing what to do is what saves the day. The author created a "translation layer" that turns raw weather numbers into simple traffic-light signals: Green (Low Risk), Yellow (Moderate Risk), and Red (High Risk).

They tested this on three major Colombian agricultural zones: the coffee-growing highlands, the rice-producing lowlands, and the flower-growing regions.

For Coffee: The system found that temperature matters a lot. A small shift in the average temperature was enough to trigger a warning.
For Flowers: These crops are incredibly sensitive to temperature. The AI found a strong link (a correlation of 0.66) between temperature changes and flower yields, meaning even a tiny heatwave could be a disaster.
For Rice: Interestingly, the system found a very weak link (a correlation of 0.09) between monthly rain totals and rice production. This suggests that for rice, just knowing the total rain for the month isn't enough; the timing of the rain during specific growth stages matters more, a nuance the simple monthly data couldn't fully capture.

Based on these findings, the system generates specific "decision signals." For example, in the coffee regions, if the system predicts a cumulative rainfall deficit of more than 44.4 mm over 48 hours, or a temperature anomaly greater than 2.0°C, it flashes a "High Risk" red light. This tells supply chain managers to immediately activate contingency plans, like rerouting trucks or pre-positioning inventory. If the deficit is between 29.6 mm and 44.4 mm, it's a "Moderate" yellow light, suggesting they should just increase monitoring.

The Verdict: A Promising Prototype

The author is careful to state that this is a proof-of-concept. They built the system using "synthetic" data—computer-generated numbers that mimic real Colombian weather patterns and official agricultural records from 2017 to 2024. They did not test this on real-time data from actual farmers yet, nor did they ask business owners if the advice was actually helpful.

However, the results suggest that the idea is solid. The framework successfully demonstrated that you can take ground-level weather data, run it through a smart AI model, and turn it into a clear, three-step warning system that gives supply chains about 38 hours of lead time to react. The author concludes that while the system needs real-world testing and refinement (especially for crops like rice that need more detailed water data), it offers a viable, low-cost path to making food supply chains more resilient against the whims of the weather. It's a digital shield, built from simple data, ready to protect the harvest.

Technical Summary: Real-Time Climate Risk Assessment for Supply Chain Resilience

Problem Statement
Climate variability in Colombia, characterized by irregular rainfall, temperature fluctuations, and recurrent extreme events, poses a significant risk to agricultural production and supply chain stability. While existing approaches often rely on seasonal forecasting or retrospective analysis, supply chain stakeholders require short-term, operationally actionable information (hours to days) to manage inventory, sourcing, and logistics. A critical gap exists in the literature: few frameworks integrate real-time climate nowcasting with supply chain decision-making, particularly in developing country contexts where satellite-based remote sensing infrastructure is limited or inaccessible. Current systems often fail to bridge the divide between meteorological forecasting and the specific, time-sensitive needs of agricultural supply chains using only ground-based data.

Methodology
The paper proposes a three-stage methodological framework designed to operate with data sources readily available in Colombia, specifically meteorological observations from IDEAM (Instituto de Hidrología, Meteorología y Estudios Ambientales) and official agricultural statistics from the Ministry of Agriculture (AGRONET). The framework avoids reliance on satellite imagery or computer vision.

Data Preprocessing: The system integrates daily meteorological time series (precipitation, temperature, humidity, solar radiation, vapor pressure deficit) and regional yield statistics. Data undergoes quality control (outlier detection, imputation) and normalization. Standardized Precipitation Index (SPI) and temperature anomalies are computed to capture slow-onset variations.
Climate Nowcasting Model: A supervised time-series prediction problem is formulated using a Long Short-Term Memory (LSTM) recurrent neural network. The model utilizes a sliding window of past observations (24–48 hours) to predict precipitation and temperature conditions 6, 12, 24, and 48 hours ahead. The architecture consists of two stacked LSTM layers (64 and 32 units) with dropout regularization, optimized via Adam with Mean Squared Error loss. A time-aware cross-validation scheme ensures temporal integrity, partitioning data into training (2017–2021), validation (2022), and test (2023–2024) sets.
Climate-Agriculture Impact Mapping: Empirical relationships are established by computing correlations between lagged meteorological variables and yield anomalies. Risk thresholds are defined using historical quantiles (33rd and 67th percentiles) of meteorological anomalies, categorized into Low, Moderate, and High risk levels. These thresholds are crop-specific (e.g., coffee, rice, flowers) and region-specific, translating raw forecasts into categorical risk signals.
Prototype Implementation: The framework was implemented in a controlled Google Colab environment using Python (Pandas, Keras/TensorFlow, Scikit-learn). Validation utilized synthetic historical scenarios calibrated on documented Colombian climate patterns and official statistics, rather than real-time field deployment.

Key Results
The prototype validation yielded the following findings:

Nowcasting Performance: The LSTM model demonstrated consistent performance across forecast horizons. Mean Absolute Error (MAE) ranged from 0.58 to 0.60 mm, and Root Mean Square Error (RMSE) ranged from 0.73 to 0.75 mm. Notably, F1-scores for extreme event detection improved at longer horizons, increasing from 0.51 at 6 hours to 0.66 at 48 hours, suggesting the model effectively captures underlying temporal patterns in extreme events.
Climate-Agriculture Correlations: The analysis confirmed heterogeneous sensitivities across crops. Flower yields showed a strong positive correlation with mean temperature ( $r = 0.66$ ), while coffee yields showed a moderate correlation ( $r = 0.38$ ). Rice production showed a weak correlation with monthly precipitation ( $r = 0.09$ ), indicating that aggregate monthly data may not fully capture the acute sensitivity of rice to intra-seasonal water stress.
Risk Signal Generation: The framework successfully generated actionable risk categories. For the Andean Coffee Belt, thresholds were established for cumulative rainfall deficits (e.g., >44.4 mm deficit over 48 hours indicating "High" risk) and temperature anomalies (>2.0°C). These signals provide a mean lead time of approximately 38 hours for high-risk alerts, sufficient for supply chain adjustments.

Significance and Claims
The paper claims to provide a structured, methodological foundation for integrating short-term climate nowcasting with supply chain risk assessment in developing country contexts. Its primary contributions are:

Operational Viability without Remote Sensing: It demonstrates that effective nowcasting and risk mapping can be achieved using only ground-based meteorological stations and official administrative statistics, circumventing the data access and infrastructure constraints typical of satellite-dependent systems.
Bridging the Forecast-Decision Gap: The framework explicitly translates atmospheric dynamics into supply chain decision signals (inventory pre-positioning, sourcing diversification, transport rerouting) through a modular, threshold-based architecture.
Proof of Concept: The prototype validates the logical consistency and technical feasibility of the approach, establishing a template for future implementation using national datasets from IDEAM and the Ministry of Agriculture.

Limitations and Future Work
The author maintains a modest stance regarding the current scope. The study is a proof-of-concept utilizing synthetic data calibrated to historical patterns; it does not yet incorporate real-time operational data, stakeholder feedback, or actual supply chain outcome tracking. The framework currently addresses only meteorological drivers, excluding pests, soil conditions, and management practices. Future work prioritizes the transition to operational deployment, the integration of real-time IDEAM and AGRONET data, stakeholder engagement to refine thresholds, and field validation to establish causal links between risk signals and supply chain resilience.

Technical Summary: Real-Time Climate Risk Assessment for Supply Chain Resilience

More like this