Land cover classification from satellite images is one of the main Earth observation (EO) tasks starting from the launch of first EO satellites. It is used for producing land cover maps showing areas covered by forests, agriculture areas, water basins, roads, buildings and other categories that are needed for the accounting of land resources. We have created an automatic classification technology which is unsupervised and highly robust (hereafter Dynland). Our working prototype allows users to classify multispectral geospatial images into land use and land cover classes. Dynland technology enables the use of machine learning methods in applications where it was not possible before due to the lack of sufficient training data. For others, lesser requirements on training data mean significantly lower cost and faster delivery of geospatial analytical results.

Dynland unsupervised classification technology for geospatial image analysis is on the overlap of two industries: Earth observation (EO) and machine learning (ML). EO capabilities are rising with 596 satellites devoted to Earth observation. EO data has become widely available through public investment in EU (Copernicus programme) and private investment in the USA (190 satellites in Planet Labs constellation alone). Daily optical imaging anywhere in the world is available. EO Data acquisition through unmanned aerial vehicles (UAV) like drones has become mainstream. UAV based EO data is already used in agriculture and will enter other industries with an integration of UAVs into shared airspace.

We have done an internal and third-party comparison of our clustering algorithm with most popular clustering and classification algorithms. Tests were done on both synthetic as well as real multispectral satellite data. Our algorithm outperformed other clustering algorithms in terms of accuracy of making proper clusters of spectrally close pixels related to one subclass of the data. Classification with Dynland is possible in circumstances where supervised algorithms fail due to a lack of sufficient training data.

Interviews with potential users of EO data reveal problems that we see as opportunities:

  • collection of reliable training data in many applications is not only expensive but often impossible
  • ground truth data is not shared among projects even in cases where it would add value
  • building internal competence, purchasing expensive remote sensing software packages, hardware and training often outweigh the benefits
  • there are inefficiencies in commercial EO satellite data pricing – minimum order sizes of 25km2 to 100km2 at 5-100 EUR/km2 is often too much for small, geographically or temporally scattered EO tasks.

We have started commercialization of Dynland that address the identified market problems. Our technology is generic enough to add value in most geospatial image classification tasks. Our priority applications are in classifying multispectral imagery in forestry, precise agriculture, insurance and land cover classification for public institutions. There is a significant demand for fully automated value-added services that do not require training data. Fully automated EO services are among the fastest-growing segments of EO industry with 38% CAGR.

Our technology provides a highly accurate classification of the objects on the image without the use of training data. It takes a satellite image (1), forms clusters of pixels, then assigns classes to clusters (2) and provides users with information about classes and their sizes (3).

For a more technical description of the technology, please read our public whitepaper.

Information about the completed and ongoing projects related to Earth Observation can be found here.