Watching Trade from Space: Nowcasting and Spatial Extrapolation of Port-Level Maritime Trade Using Satellite Imagery

This paper presents a novel framework that combines synthetic aperture radar imagery, nighttime lights, and port characteristics to accurately nowcast and extrapolate monthly port-level maritime trade, demonstrating its effectiveness in detecting trade shifts—such as Russia's reorientation after 2022 sanctions—while remaining robust to signal manipulation where traditional AIS methods may fail.

The Gist

Imagine you are trying to guess how busy a specific grocery store is, but the store manager refuses to show you the sales receipts, and the security cameras are broken. How would you know if they are selling more or less than last month?

You might look out the window. You'd count how many cars are in the parking lot, see if the lights are on late at night, and notice if the delivery trucks are coming and going. Even without the official receipts, these visual clues tell you a lot about the store's activity.

This paper does exactly that, but for shipping ports around the world.

Here is the story of the paper, broken down into simple concepts:

1. The Problem: The "Black Box" of Trade

Usually, when a country ships goods, they keep official records (like a ledger). But sometimes, countries hide these records.

• Why? Maybe they are under sanctions (punishments), maybe they are trying to hide illegal trade, or maybe their government just isn't good at keeping track.
• The AIS Problem: Ships usually have a "transponder" (like a GPS tracker) that tells the world where they are. But if a ship is doing something shady, they can turn that tracker off. It's like a car driving with its headlights off; you can't see it on the map.

2. The Solution: "Watching from Space"

The author, Yonggeun Jung, decided to stop looking at the ships' GPS and start looking at the port itself from space. He used three types of "eyes" in the sky:

• The "Radar Eye" (SAR): This is a special camera that works day and night, even through clouds. It bounces radar off the ground.
  • The Trick: If a ship moves, the radar image changes. If a container is stacked up, the radar bounces back differently. It's like noticing that the furniture in a room has been rearranged, even if you can't see the people moving it.
• The "Night Light Eye" (Nighttime Lights): Ports work 24/7. If a port is busy, it's bright at night. If it's quiet, it's dark. The author measured how bright the port was and how much that brightness changed day-to-day.
• The "ID Card" (Port Characteristics): Not all ports are the same. Some are huge with deep water; some are small. The author fed the computer a list of facts about each port (like its depth and size) so the computer knew the "personality" of the port.

3. The Brain: The Machine Learning Detective

The author didn't just look at the pictures; he taught a computer program (called XGBoost) to be a detective.

• The Training: He showed the computer data from U.S. ports where he did have the official receipts. He said, "Look, when the radar shows this much movement and the lights are this bright, the trade value was this much."
• The Result: The computer learned the pattern perfectly. It could predict the trade numbers with 94% accuracy just by looking at the space photos.

4. The Big Challenge: Can it work elsewhere? (The "Hawaii Test")

The computer was trained on U.S. mainland ports. Could it guess what was happening in Hawaii (which is far away and looks different)?

• The Failure: When the computer tried to guess the exact dollar amount, it failed. It's like trying to guess the price of a house in Hawaii based on a house in New York. The "base price" is just too different.
• The Fix (Anchoring): The author realized the computer is great at seeing changes, but bad at guessing absolute numbers.
  • The Analogy: Imagine you are guessing the temperature in a new city. You don't know the exact temperature, but you know it's 5 degrees warmer than yesterday. If you know yesterday's temperature, you can guess today's.
  • The author used a "calibration trick" (anchoring) to fix the base level. Once he did that, the predictions were spot on.
• The Golden Rule: The computer is amazing at telling you if trade is going up or down (percentage change), even if it can't tell you the exact dollar amount without a little help.

5. The Real-World Test: Russia After Sanctions

In 2022, Russia was sanctioned, and they stopped reporting their trade data. They also turned off many ship trackers to hide their oil sales.

• The Experiment: The author applied his "Space Eye" model to Russian ports.
• The Discovery:
  • European Ports: The model saw activity dropping. (The lights went dim, the radar showed less movement).
  • Far East Ports: The model saw activity spiking.
• The Conclusion: This confirmed that Russia was secretly rerouting its oil and goods from Europe to Asia (China, etc.) using "dark ships" (ships with trackers off). The space camera saw the physical movement even though the official reports were silent.

The Takeaway

This paper proves that you don't need secret government data to know what's happening in global trade. By combining satellite radar, nighttime lights, and smart computer learning, we can "see" trade activity even when countries try to hide it.

It's like having a superpower: You can watch the world's economy move, even when the players try to play in the dark.

Technical Summary

1. Problem Statement

Official international trade statistics are frequently delayed, incomplete, or strategically suppressed, particularly in environments involving sanctions, weak statistical capacity, or conflict. While Automatic Identification System (AIS) data has been used for trade nowcasting, it is vulnerable to strategic manipulation (e.g., vessels disabling transponders to evade sanctions or hide illicit trade).
The Gap: There is a lack of high-frequency, reliable methods to measure port-level maritime trade using only publicly available data that does not rely on voluntary signal transmission from vessels.

2. Methodology

The author proposes a machine learning framework that combines remote sensing data with static port characteristics to predict monthly trade values and weights.

A. Data Sources

The model relies exclusively on publicly available data:

1. Synthetic Aperture Radar (SAR) Imagery (Sentinel-1):
   • VV Polarization (Movement Proxy): Calculates the absolute pixel-wise difference between consecutive images to detect changes in port activity (ship movements, loading/unloading).
   • VH Polarization (Stock Proxy): Measures mean backscatter intensity to detect metallic objects (containers, cranes, vessels), serving as a proxy for cargo stock and infrastructure.
2. Nighttime Lights (NTL) (VIIRS):
   • Mean Intensity: Proxy for average operational intensity.
   • Standard Deviation: Captures short-run fluctuations in daily operations.
   • Light Area Ratio: The share of illuminated pixels above a threshold, proxying the spatial footprint of port infrastructure.
3. Port Characteristics (World Port Index - WPI):
   • Includes 91 static attributes (e.g., channel depth, berth dimensions, cargo handling facilities) to control for cross-port heterogeneity.
4. Ground Truth: Monthly U.S. port-level trade data (value and weight) from the U.S. Census Bureau (2016–2024).

B. Model Architecture

• Algorithm: XGBoost (Extreme Gradient Boosting), chosen for its ability to handle tabular data, moderate sample sizes, and capture non-linear relationships.
• Training Scope: 64 U.S. ports (classified as "Small" or larger) from Jan 2016 to May 2022.
• Testing Scope: June 2022 to Dec 2024.
• Specifications:
  1. Full Model: Satellite variables + Port characteristics.
  2. Satellite-Only: Satellite variables only.
  3. Port-Only: Static characteristics only.

3. Key Contributions

1. Novel Data Fusion: Demonstrates that combining SAR (movement/stock) and NTL (operational intensity) with static port attributes can accurately nowcast trade without AIS data.
2. Robustness to Signal Manipulation: Provides a monitoring tool that remains effective even when vessels intentionally disable transponders (a critical advantage over AIS-based methods in sanction regimes).
3. Spatial Extrapolation Theory: Formalizes the conditions under which spatial machine learning succeeds or fails. The paper proves that while absolute levels are difficult to extrapolate to unseen regions, percentage changes (growth rates) can be recovered with high accuracy.
4. Anchoring Strategy: Proposes a method to correct level predictions in new regions by using a single observation to align the predicted intercept, significantly improving spatial generalization.

4. Key Results

A. Nowcasting Performance (U.S. Ports)

The model achieves strong out-of-sample accuracy:

• Trade Value: $R^2 = 0.945$ (Full Model), $R^2 = 0.863$ (Satellite-Only).
• Trade Weight: $R^2 = 0.880$ (Full Model), $R^2 = 0.776$ (Satellite-Only).
• Feature Importance:
  • In the Full Model, static port attributes (e.g., oil terminal depth) dominate level prediction.
  • In the Satellite-Only model, VV image differences (31%) and NTL mean (25%) are the top predictors.
  • Complementarity: Static attributes pin down cross-sectional levels, while satellite signals capture within-port temporal dynamics.
• Placebo Test: Shuffling satellite variables results in $R^2 \approx -0.02$, confirming the model relies on meaningful signals, not temporal persistence.

B. Spatial Extrapolation (Hawaii)

• Challenge: Predicting Hawaiian ports using a model trained only on the U.S. mainland resulted in negative $R^2$ for absolute levels due to unobserved geographic/institutional differences.
• Solution (Anchoring): By applying a constant offset (anchoring) based on the first month's observation, the model recovered high accuracy ($R^2 = 0.874$ for value).
• Insight: Satellite-only models extrapolated better than models with categorical port characteristics, suggesting high-dimensional static features can hinder generalization to disjoint spatial domains.

C. Application to Russia (Post-2022 Sanctions)

• Method: Applied the framework to 15 Russian ports, focusing on percentage changes rather than absolute levels to avoid anchoring errors.
• Findings:
  • European Ports: Predicted declines (e.g., Kaliningrad -1.44%), consistent with trade disruption.
  • Far East Ports: Predicted increases (e.g., Korsakov +0.43%, De Kastri +1.58%), indicating a reorientation of trade toward Asia.
  • Black Sea: Predicted increases in military/logistics hubs (Novorossiysk, Rostov-on-Don), likely capturing non-commercial military activity rather than commercial trade.
• Significance: The framework successfully detected trade reorientation and "dark shipping" behaviors despite the absence of official Russian trade data and potential AIS signal suppression.

5. Significance and Limitations

Significance:

• Policy Relevance: Offers a transparent, reproducible tool for monitoring trade in sanctioned or data-poor regions.
• Methodological Advance: Clarifies the "spatial extrapolation" problem in ML, showing that while levels are hard to predict across domains, changes are robustly recoverable.
• Complementarity: Fills the gap left by AIS-based methods when signals are manipulated.

Limitations & Future Work:

• Granularity: Currently operates at the aggregate port level; cannot distinguish bilateral flows or specific commodity types.
• Non-Commercial Activity: Cannot distinguish between commercial trade and military/logistics activity (as seen in the Black Sea results).
• Data Scope: Relies on U.S. training data; future work requires expanding training to diverse global contexts to improve global applicability.
• Causality: Designed for measurement/nowcasting, not causal inference.

Conclusion

Jung's paper establishes a robust, publicly accessible framework for monitoring global maritime trade. By leveraging the physical footprints of port activity visible from space, it overcomes the limitations of traditional statistics and vulnerable vessel tracking systems, providing a critical tool for economic analysis in times of geopolitical instability.