Night Lights is a type of imagery collected at night that can help identify vessels using lights. Light from small boats as well as large vessels is detected based on the brightness of those lights. The technical term for this type of data is Visible Infrared Imaging Radiometer Suite, often referred to as VIIRS.

How it works: Night Lights imagery is a low resolution electro-optical imaging system that capture images of the earth at night. It’s possible to detect vessels from the light they emit at night. The sensors have high spectral fidelity, meaning they can differentiate wavelengths across the electromagnetic spectrum.

While the satellites can detect typical deck or running lights on vessels, the brightest and most reliable detections tend to be from the working deck lights or the lights used to lure plankton or squid in industrial fishing fleets.

Value / Challenges: A key benefit of Night Lights data is that data is collected globally every night as opposed to most satellites that collect imagery during the day and typically do not cover the high seas or the outer extents of EEZs. 

The three satellites of Night Lights follow one another collecting imagery, which introduces the possibility to learn a vessels course. This also means you may see a vessel detected multiple times within a few hours. 

Finally, has lowest latency of all the regularly available data processed by Skylight at around 2-2.5 hours.

The challenges of night lights are two fold. The first is that not all vessels operate at night with bright enough lights to detect. The second is that environmental challenges reduce the amount of usable imagery or can cause false positives. A later section explores these challenges more in-depth.

Satellites and resolution

Night lights detections come from the three satellites from US National Oceanic and Atmospheric Administration (NOAA) and US National Aeronautics and Space Administration (NASA).  

Source

Three satellites: SUOMI-NPP and NOAA-20 (JPSS-1) and NOAA -21 (JPSS-2). These satellites all have similar sensors. The addition of JPSS-3 and JPSS-4 satellites are scheduled for the coming years.

Key stats

• Coverage: Global by each satellite
• Resolution: 750 meters
• Latency: 2-2.5 hours
• Revisit rate: Daily

Coverage

Each of Night Lights satellites images the entire world every night between the hours of 1-4 am local time. 

The frames for Night Lights are not shown in Skylight because they can be assumed to be global and would otherwise overwhelm the screen with frame. Each swath (width of area in each image) is 3000km.

Resolution

The resolution of Night Lights imagery is 750 meters (each pixel is 750 meters x 750 meters). 

The value of such low resolution imagery is in the vast coverage resulting in daily, global coverage. However, it comes with challenges. With each pixel representing 562,500 sq meters (0.56 sq km). This results in single pixel detections potentially representing more than one vessel if, for example, two vessels were alongside one another.

Latency 

The average latency (delay from time of imaging) for Night Lights is 2-3 hours. 

This range includes the time for the satellite to capture the image, send the image to earth, and then for Skylight to process the data. The time to send the image from the satellite to earth accounts for a majority of the latency.

The chart below is a 2 week sample. Note that the day-to-day average latency is mostly between 2-3 hours, but sometimes as short as 90 minutes and sometimes longer.

Challenges and limitations

Using Night Lights data to detect vessels has historically been challenged by a number of environmental factors. The Skylight team has been able to resolve or limit many of the common issues, except for situations of strong moonlight reflecting off of clouds (glint).

Moonlight + clouds

Moonlight reflecting off clouds, otherwise known as glint, is the biggest challenge preventing detections from Night Lights data.   

In such conditions, Skylight attempts to omit such false positive detections. Analysts should exercise additional discretion during periods around a full moon.

Skylight is able to detect vessels under clouds when moonlight is not a concern. The detection may appear larger due to the cloud diffusing the light. 

Model details

The modeling strategy adopted a three stage approach. The first stage consisted of a classical computer vision model, trained without supervision, to extract all possible sources of light. This was achieved with a simple 2D kernel. 

In the second stage, all known non-vessel light sources (lightning, gas flares, moonlit clouds, the northern and southern lights, and ionospheric particles from within the South Atlantic Anomaly) are removed through a series of post-processing steps. These non-vessel light sources often exhibit stereotyped distributional patterns (unlike vessels) that is amenable to rules based logic. Additionally, we implemented statistical tests to identify unusually geographically distributed vessels coincident with scan lines and suppressed false positives at the frame’s extremities due to the “noise smile" to control the false positive rate. 

The final stage involved filtering all positive detections through a regularly updated 2d CNN. This CNN was trained on human annotated image labels (correct/not) with four channels (nano watts, land water masks, moonlight, and clouds [11, 12]). This model was specifically designed to run in resource constrained environments, requiring only modest hardware (2 GB RAM, no GPU).

End-to-end deep learning based approaches were also evaluated. However, given the simplicity and limited spatial extent of the objects, and their sparse distribution, end-to-end deep learning based models required significantly more computational power to achieve performance comparable to our hybrid design. Our hybrid approach is designed to be highly efficient, which allows for regular and economical retraining using new labeled data. Such efficiency is particularly beneficial for continuoul improvement, ML specific continuous integration and continuous delivery (CI/CD) pipelines and model-specific unit/integration/regression testing. Thorough testing (especially within the CICD framework) is beneficial for preventing regressions during phases of fast paced development, and is typically prohibitively expensive with conventional large-scale DNNs.

Overcoming challenges

During the development of the Machine Learning model, the team overcame several challenges that could otherwise cause false positives.

In addition to the examples below, the Skylight team also mitigates against any false positives from imaging artifacts, aurora, and lights from land.

Lightning 

The brightness of lightning strikes create a common horizontal stripe in the image. The machine learning model is able to recognize such examples as candidates for a vessel detection and remove them.

Oil platform gas flares

Gas flares may look like vessels to the naked eye. However, the light emitted from gas flares is visible within a particular spectral band (M10) that lights from a vessel are not. The Skylight team is able to identify these instances and remove them from candidates for a vessel detection

South Atlantic Anomaly

Ionospheric noise particles hit the Night Lights/VIIRS sensor causing the appearance of a lighted vessel. The model is able to recognize the extreme brightness of these signals to remove them via a noise filtering technique (erosion>dilation). 

Additional resources for in-depth information on Night Lights/VIIRS.

• Data documentation
• Black Marble User Guide - This was consulted extensively by Skylight Machine Learning team as the VIIRS model was developed
• Correlation of vessel category and vessel brightness
• Cloud detection (SSEC docs)
• Land sea masking (NASA docs)
• Night Lights Open sourced code/model (in Skylight’s public codebase)