Can Fires, Night Lights, and Mobile Phones reveal behavioral fingerprints useful for Development?

Fires, lights at night and mobile phone activity have been separately used as proxy indicators of human activity with high potential for measuring human development. In this preliminary report, we develop some tools and methodologies to identify and visualize relations among remote sensing datasets containing fires and night lights information with mobile phone activity in Cote D’Ivoire from December 2011 to April 2012.

💡 Research Summary

This paper investigates whether three ubiquitous digital signals—satellite‑detected fires, night‑time lights, and mobile phone call detail records (CDRs)—can be combined to produce behavioral “fingerprints” useful for development monitoring. The authors focus on Côte d’Ivoire, using data from December 2011 to April 2012. Fire locations are obtained from NASA’s MODIS‑based FIRMS product (≈59 000 active fire detections). Night‑time illumination is derived from VIIRS Day‑Night Band composites at 3 km resolution. Mobile phone data come from Orange’s D4D challenge: (a) hourly aggregated call volumes between 1 231 antenna pairs, and (b) high‑resolution trajectories of 50 000 randomly sampled users, with timestamps rounded to the minute.

The methodology proceeds in four steps. First, fire points are spatially matched to antenna locations within a 1 km radius, identifying 95 antennas associated with 109 fire events. Second, the authors extract the average VIIRS radiance within a 7.5 km buffer around each antenna and apply k‑means clustering (k = 3) to separate the sites into three groups that correspond, after visual inspection, to large cities (8 antennas), small towns (15 antennas), and rural/road areas (72 antennas). This classification is independently validated by counting individual trajectories that pass near each antenna on the fire day; the number of trajectories correlates strongly with the light‑based urban‑rural split, confirming that night‑lights can serve as a proxy for population density in this context.

Third, the paper examines the temporal dynamics of call activity around fire events. For each antenna, hourly call counts are normalized by the antenna’s maximum hourly volume observed in the five‑day window centered on the fire (t = 0). The authors then average across antennas within each cluster. The results reveal distinct patterns: in rural areas, the usual bimodal daily peaks (morning and evening) invert after a fire—morning activity rises while evening activity falls—suggesting heightened information‑seeking behavior in the immediate aftermath. In small towns, terminating calls (calls that end) surge in the morning of the day after the fire, possibly reflecting reporting or coordination activities. In large cities, overall call volume drops on the fire day, which may reflect network congestion, displacement, or temporary work stoppages. Longer‑term analyses (weekly to monthly) are hampered by strong periodicities inherent to prepaid mobile usage (high activity at the start of the month, declines toward month‑end), holiday effects (reduced calls around Christmas), and occasional spikes on the first of the month. These confounders make it difficult to isolate fire‑related trends over longer horizons, highlighting the need for more sophisticated normalization or longer observation windows.

Finally, the authors develop an interactive visual‑analytics platform called MOBILOMICS, built with Processing and the Unfolding Maps library. The tool integrates OpenStreetMap basemaps, fire locations, antenna positions, and individual trajectories, allowing users to filter by time, space, and cluster type. Sample screenshots demonstrate how one can explore movement patterns before, during, and after a fire at multiple spatial scales, facilitating hypothesis generation and exploratory analysis.

In the discussion, the authors argue that merging fire, night‑light, and mobile phone data can improve the granularity of socioeconomic indicators traditionally derived from night‑lights alone, enabling near‑real‑time monitoring of urban versus rural activity, disaster impact, and recovery. They acknowledge several limitations: the relatively short five‑month observation period, the modest sample of 50 000 users (which may not be representative of the whole population), and the lack of ground‑truth data on fire type (urban, agricultural, or conflict‑related). They propose future work to extend the temporal span, incorporate additional data sources (e.g., census, economic surveys, emergency response logs), and refine models to distinguish fire typologies and their distinct socioeconomic effects.

Overall, the paper presents a proof‑of‑concept that integrating remote‑sensing fire and illumination data with high‑resolution mobile phone records can uncover nuanced behavioral signals relevant for development and humanitarian applications, while also outlining the methodological challenges that must be addressed to move from exploratory analysis to robust, policy‑ready indicators.