Machine Learning Offers New Opportunities for Geospatial Applications

Four Powerful ML Apps

Thanks to constellations of imaging satellites, advanced aerial cameras and scanners and various other collection devices, the volume of available geospatial data has grown beyond the capability of humans to manually manage and analyze the datasets. To leverage this abundance of information, machine learning is the new programming paradigm that effectively extracts the wealth of knowledge contained in millions of petabytes of archived and frequently refreshed data. Using machine learning, large datasets are reviewed and analyzed in a fraction of the time compared to previous methods.

Geospatial Platform in the Cloud
To expedite searching for data in areas of interest around the world and analyzing large datasets with machine learning, a company based in Berlin named UP42 has developed a geospatial platform in the cloud. UP42 fulfills the need for data and processing algorithms as well as provides the infrastructure for high-powered computing.

For customers interested in developing machine learning algorithms that solve problems or answer a specific question, UP42 provides the building blocks for powerful geospatial products. A block essentially is a ready-to-use unit of data or processing algorithm that customers string together to form workflows. The basic data handling algorithms give developers a head start with processing blocks such as “Pan-sharpening SPOT/Pléiades images” and “K-Means Clustering for unsupervised classification.” UP42 also provides access to data blocks from multiple sources ranging from 0.5m Pléiades images to Landsat-8 and NEXTMAP digital surface models and digital terrain models.

The developer platform created by UP42 is made accessible to its customers through APIs. The platform enables browsing the datasets and selecting data blocks that meet the customer’s criteria, before applying custom or off-the-shelf algorithms. Developers can choose either to put their custom algorithms onto the platform in a private block or publish a processing block that is accessible to the whole UP42 community for a fee. In addition to facilitating analysis of geospatial data with machine learning, service offerings in the cloud are scalable to meet the need for any level of computing power.
“It is an exciting time for machine learning as a huge amount of resources is being put into the development of algorithms for many industries,” says Rodrigo Almeida, UP42 data scientist. “Many are available in open source code which really encourages more creativity and development of new geospatial applications.”

Repetition Enhances Accuracy
The traditional method of programming involves writing code to give instructions for a defined task, such as 1 + 1 = 2, whereas with machine learning, the programmer builds a model that is trained to recognize a situation and return an optimal answer. By repeatedly exposing the model to similar examples, the algorithms “learn” how to return a desired result. The more data used for training, the more accurate the results produced by the algorithms, so developers need access to large quantities of images.

Off-the-shelf algorithms save developers a great deal of time. For example, UP42 partner Orbital Insights describes its Car Detection Block like this: “The training set contains 180,000 images and spans many variable conditions including time of day, time of year, terrain, configuration of vehicles, etc. This algorithm was tested by a third party using a validation set consisting of approximately 100,000 marked cars in 50 countries spanning 20,000 images of deserts, ports and parking lots across six continents.”

Rather than waste time re-doing all this careful training and testing, UP42 customers are able to focus their efforts on developing new complex algorithms and building on the foundation that has already been created.

“Many of UP42’s customers are interested in high resolution satellite imagery,” says Almeida. “Our marketplace saves time by allowing them to search for their area of interest and immediately purchase the data online. We also provide a lot of free and open source lower resolution data, such as MODIS and Landsat. With our global coverage, customers find they can develop many interesting applications by combining free low-resolution and more costly high-resolution imagery.”

Four Powerful ML Applications
For geospatial professionals, machine learning shows the most promise in at least four general categories of problem solving: object identification, monitoring, change detection and spotting trends. Algorithms are trained to conduct the tasks necessary to search the areas of interest, extract information about the target areas, and compile reports to support decision making.

Object Identification –

Machine learning is well-suited for object identification because of the consistency of shapes and characteristics of objects around the world. A passenger vehicle is easily recognizable, with some variations such as convertibles, pick-up trucks, minivans, etc. A tree always looks like a tree, with some variation between deciduous and evergreen species. The training process for machine learning involves labeling similar objects found in different countries, seasons, times of day, colors, etc., until the algorithm automatically recognizes the object a high percentage of the time.

One application for object identification is tracking ships in the ocean to detect illegal fishing or trade. Many organizations, from environmental protection agencies to insurance companies and national government authorities, have compelling reasons to keep a close watch over the open seas. Using imagery from remote sensing satellites to find and track ships is the only feasible method of covering these large areas. Airbus is one of several companies that has developed a Ship Detection algorithm trained to identify ships that meet certain size and shape requirements in SPOT satellite imagery.

Monitoring –

Monitoring large areas of land, from cities to continents, is more efficient with machine learning applied to geospatial data. Numerous applications, such as natural hazards, vegetation encroachment, and agriculture, benefit from regular status updates and early warnings to catch problems before major damage occurs. Careful, frequent monitoring is of interest to governments, commodity brokers, insurance adjusters, environmental groups, farmers and many others.

On a global scale, monitoring crops is valuable for making yield estimates and predicting shortages and food chain disruptions, while at a micro-level, an algorithm may detect water stress in one part of a field, which could point to a malfunctioning irrigation system. For example, algorithms developed by geospatial service provider Vasundharaa calculate a Soil Adjusted Vegetation Index and a Moisture Stress Index to indicate crop health based on satellite imagery.

Change Detection –

Machine learning algorithms distill knowledge from large volumes of data, and the resulting reports help visualize the answers. This is particularly valuable for tracking environmental problems, such as the loss of wetlands and deforestation. For example, on a land use classification map comparing forests year to year, a color scheme that assigns red to areas where trees have been removed and green to remaining forests quickly communicates where deforestation is taking place. The same method applies to wetlands, agricultural land, open space, etc.

The UP42 platform offers satellite imagery from multiple sources, such as the Pléiades and Sentinel constellations, that provide frequent revisits and global coverage with multispectral data useful for change detection.

Spotting Trends –

By studying remotely sensed data collected by Earth observation platforms, patterns of life over time can be accurately identified with machine learning algorithms. In turn, this information is used for predictive analytics, which influence development of transportation networks, energy infrastructure, housing supply, and many other areas of our lives.

For example, Orbital Insight’s car detection algorithm uses wide area object detection on Pléiades imagery to accurately identify and quantify cars. This saves analysts significant time when conducting pattern of life analyses and generating activity-based intelligence. Governments and businesses are interested in this type of information, as it could indicate growth or recession in the economy or an increase or decrease in population.

Leveraging Available Information
Imagine a person trying to analyze satellite imagery of the entire United States, or of the globe, to identify and evaluate economic, societal, and environmental activities. The volume of data would be overwhelming, and the lengthy process would render the results nearly useless. As described by James Crawford, CEO of Orbital Insights, at the 2018 Artificial Intelligence Conference, it would require eight million people doing nothing but looking at satellite imagery 24/7 to ensure every photo taken each day is just viewed, let alone analyzed.

Machine learning applied to geospatial data has tremendous potential as a decision-support tool in many areas, including risk mitigation, early warning, data-informed management, predictive maintenance, etc. Algorithms extract the necessary information through an automated repeatable process, faster than can possibly be accomplished by humans alone. The UP42 platform provides all of the necessary building blocks — access to multi-source data, off-the-shelf algorithms and cloud computing power — for customers to tap into timely knowledge on a global scale.

Category: Education

This entry was posted on Wednesday, December 9th, 2020 at 2:51 pm. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.