Since its introduction, satellite-imaging technology has been driven by the goal of providing consistent monitoring of Earth’s environment and human activities. Benefiting from the recent deployment of numerous civilian observation satellites into orbit, the volume of accessible satellite imagery and its derived data has experienced explosive growth, enhancing capabilities for observing human activities and measuring commercial behaviors. Recurring satellite imagery refers to data collected by satellites that periodically observe the same geographic area over time. These satellites typically feature lower spatial resolution but larger coverage areas, making them suitable for monitoring extensive land surfaces. In contrast, tasking satellite imagery involves customized, on-demand services for single-pass image acquisition of specific areas, generally producing higher-resolution data at a greater cost. With the rapid advancement of machine learning in recent years, low-cost recurring satellite imagery, augmented by machine learning techniques, can now generate diverse commercial intelligence to support business decision-making.

Growing Attention to Recurring Satellite Imagery in Practice

Due to its periodic observation capability, recurring satellite imagery is used to monitor traffic, storage levels of oil and gas resources, and track volume and pile morphology changes over time for specific materials (like coal or waste) in storage areas. Simultaneously, terrain alterations resulting from human activities serve as significant observation targets, such as construction progress at sites, crop harvest status, wildfire propagation area, refugee camp populations, and regions affected by natural disasters. The periodic, frequent, and comprehensive Earth observation provided by recurring satellites yields massive image datasets, while deep learning classification algorithms can efficiently perform tasks like identification, categorization, and dynamic change detection. The resulting analytics help professionals analyze market conditions more effectively. Platforms integrating satellite technology with artificial intelligence can closely track inventory levels and transportation flows of various bulk commodities, offering insights into macroeconomic and microeconomic dynamics from a time-series perspective. Leveraging these advantages, numerous data service firms worldwide now offer cargo flow monitoring services based on AI-enhanced recurring satellite imagery. For example, “MySteel” (https://www.mysteel.net/), a leading global data service provider for bulk commodities and related industries in mainland China, provides clients with maritime cargo flow monitoring data within China.

Practical Applications of Recurring Satellite Imagery

• Industrial Activity Monitoring: The potential commercial applications of AI-analyzed recurring imagery are vast. For instance, time-series data from Digital Surface Models (DSM, a 3D computer graphics representation of terrain) can effectively monitor industrial activities, such as inventory levels of goods in a specific area. This enables professionals, assisted by AI feature recognition, to estimate volume, height, and other metrics of target objects over time. Such data allows for real-time monitoring of supply/consumption dynamics, facilitating better arbitrage decisions. For example, to understand inventory fluctuations over the past six months at the Richards Bay Coal Terminal (RBCT) in South Africa, a large coal storage facility covering 1.6 square kilometers, one could use recurring imagery from satellites like SkySat to generate DSMs. If 40 irregularly spaced images are available over these six months, the resulting DSMs can represent the daily above-ground height of coal piles, convertible into inventory volume. As illustrated in a hypothetical figure (as shown below), darker colors could indicate greater pile height. This information easily yields metrics like average inventory, min/max levels, and average drawdown rate, aiding in predicting futures physical delivery prices, spot purchase prices, and subsequent product pricing.

• Oil Storage Measurement: This process typically begins by identifying storage tanks using optical imagery, then processing Synthetic Aperture Radar (SAR) data. Through steps like localization, registration, and feature extraction via deep learning algorithms, the unique reflection signatures of tanks in SAR imagery determine their location and state (e.g., floating roof position). Finally, the liquid volume inside the tanks is calculated, providing real-time, independent global data for oil market analysis. Taking the United States as an example, its crude oil inventories are influenced by OPEC production decisions, domestic political events, tax changes, and other factors. Inventory levels affect domestic crude prices, with higher stocks generally leading to lower prices. As the world’s top oil producer in 2023 and second-largest importer in 2022, the US has numerous oil import/export enterprises. Estimating US domestic oil inventories via satellite/AI and comparing them with historical data can indicate whether current prices are relatively high or low, informing decisions on going long/short on relevant import/export companies for arbitrage.

• Vehicle Detection and Economic Activity Assessment: On higher-resolution imagery (e.g., from Pleiades, WorldView), deep learning algorithms can detect vehicles. Despite challenges like high dataset annotation costs, need for diverse data, and algorithm selection difficulties, this can estimate factory output by monitoring parking lot vehicle counts near plants or assess popularity of commercial districts/activity in office areas. Medium-resolution PlanetScope imagery can estimate parking occupancy by analyzing pixel density features in parking areas using models like Gaussian Mixture Models. This information greatly assists professionals in commercial planning and corporate research: determining area popularity aids brands in site selection decisions, while accurate patronage estimates combined with average transaction values can yield unofficial sales data. Comparing this with official figures may help assess potential financial misreporting, informing decisions on shorting a company’s stock.

Prospects and Challenges for Recurring Satellite Imagery

The application of recurring satellite imagery combined with deep learning in commerce holds many unexplored possibilities and significant untapped potential. For instance, AI algorithms could analyze low-resolution recurring imagery to inexpensively count idle vessels in ports, estimate their draft, tonnage, etc., providing insights into specific route capacity utilization, reflecting charter market supply/demand dynamics, and aiding chartering and financial decisions. Essentially, any task requiring object identification while preserving time-series information could be supported by this combination. However, further development is needed regarding annotation costs, data diversity, and algorithm accuracy. Nevertheless, the evolution of recurring satellite imagery and AI holds immense potential for enhancing corporate decision-making and identifying arbitrage opportunities. By continuously refining models and algorithms to improve the accuracy and efficiency of feature extraction and classification, AI-integrated recurring satellite imagery can empower businesses to remain highly responsive to market changes, consistently staying a step ahead.

Reference

[1] Franchis, C., Drouyer, S., Facciolo, G., von Gioi, R. G., Hessel, C., & Morel, J-M. (2023).”The Exploitation of Recurrent Satellite Imaging for the Fine-Scale Observation of Human Activity.”. Machine Learning and Data Sciences for Financial Markets.

The work described in this article was supported by InnoHK initiative, The Government of the HKSAR, and Laboratory for AI-Powered Financial Technologies (AIFT).
(AIFT strives but cannot guarantee the accuracy and reliability of the content, and will not be responsible for any loss or damage caused by any inaccuracy or omission.)