Day 4

Detailed paper information

Back to list

Paper title FieldFinder: instance segmentation of agricultural fields in high resolution optical satellite imagery
Authors
  1. Liam Harris Airbus Speaker
  2. Stephen Hayward Airbus Defence and Space
  3. Hannah McNally Airbus
  4. Thomas Harling Airbus Defence and Space
  5. Andrew Tewkesbury Airbus
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Airbus Intelligence UK, in partnership with agrifood data marketplace Agrimetrics, has developed FieldFinder, a computer vision analytics service that uses state of the art artificial intelligence to automatically delineate agricultural fields visible in optical satellite images. Using high-resolution imagery, growers, agribusinesses, retailers and institutions can be quickly and cost effectively provided with up-to-date field boundaries at any geographic scale. Here we explore how FieldFinder uses deep learning instance segmentation to extract field polygons from images captured by Airbus’ SPOT, Vision 1 and Pléiades satellites on demand.
Traditional field boundary capture methods, such as ground surveying or digitisation using aerial photography, can be exceptionally time consuming and therefore expensive to perform. FieldFinder produces agricultural field polygons quickly and remotely using cloud computing resources, removing the inefficiencies associated with manual field boundary data capture.
Furthermore, scaling up some traditional methods over particularly large areas can be a prohibitively expensive and elongated exercise. FieldFinder provides consistent, good quality field boundaries at any spatial scale with the same high level of accuracy throughout. FieldFinder delineates boundaries using high resolution satellite imagery, providing a reliable source of information, depicting even very small agricultural fields.
At the current stage in the development of FieldFinder, several geographically specific algorithms have been trained, including those for Western Europe, Iowa (also applicable to many other parts of the USA) and Kenya (also applicable to other regions with prevalent small holder agriculture). Although the ultimate goal is to develop a single algorithm that can be deployed anywhere in the world, it is important to approach this methodically, training and validating algorithms by territory, as there can be considerable observable differences in agricultural style between territories. The current algorithms have been developed by curating spatially and temporally varied ground truth datasets from a wide selection of high resolution satellite images, ensuring a high level of accuracy and accounting for different geographic regions that demonstrate distinct features.
A number of different sources of variation are represented in the training data, including different stages in the growing season, all possible land cover types and a wide range of observable features (including non agricultural features, which must be seen by a training algorithm to reduce false detections). Data augmentation was used to further expand the available training data, incorporating possible random variation. Such data curation efforts ensured the production of good quality training data, maximising the performance of any algorithm trained, however this is also a continuous process that develops as FieldFinder is used, constantly improving the training data and therefore the algorithms.
Not only is FieldFinder always improving in terms of its performance and geographic scope, but its capabilities are also constantly being evolved, and these evolutions will also be presented. Recent work has focused on performing automatic agricultural field change detection, highlighting only those fields that have undergone observable boundary changes from one image epoch to the next. This is extremely valuable for organisations tasked with maintaining regularly updated agricultural field databases, as such a tool can significantly reduce the time and therefore cost required to update these databases. There is also ongoing research into transitioning to self supervised learning, which is a highly cutting edge paradigm for training neural networks with small amounts of training data. Data availability is often the primary blocker for the creation of Earth observation analytical algorithms, so this will not only accelerate the rollout of FieldFinder to new territories and use cases, but will benefit future algorithm development.
The computer vision and deep learning techniques employed to develop FieldFinder are evolving at a sometimes startling pace, constantly giving rise to new technologies and therefore possibilities. These techniques are powerful, can provide solutions to numerous challenges and are applicable to almost every industry that makes use of Earth observation data. Similar algorithms can be developed for the detection, classification and tracking of any kind of object of interest, to provide advanced automatic mapping capabilities, site monitoring and alerting, or even for prediction and forecasting. Airbus continues to develop these technologies, constantly furthering and enhancing the actionable intelligence that can be extracted from high resolution satellite imagery.