335

ID:

Background
Each minute 40 football fields worth of forests are lost to deforestation, usually converted to other land uses, especially commodities production. Roads and other infrastructure surfacing in forested areas are often a precursor to larger deforestation events so detecting these features can be used for early detection of deforestation to follow. Planet manages the largest fleet of earth observation satellites in the world capturing Earth's land mass just about every day at 3-4 meter resolution - a dataset uniquely suited to capturing this development.
Objective
Reduce the manual effort required to find areas of new development within the country of Brazil so that deforestation events can be tracked and potentially stopped.
Method
First, we trained an image classifier to identify roads and buildings in Planet imagery. Using hand-labeled imagery comprised of 4-band surface reflectance scenes projected to a UTM grid, we were able to train a high performing segmentation model that classified each pixel within the image as a road, building or other. We then applied this model to our near-daily planetscope imagery to extract these features within each published image. By aggregating these results on a weekly basis, we are able to remove most noise within the imagery that may be caused by shadows, clouds or other artifacts to get an accurate representation for that given week. Using these weekly layers, we are able to detect when a pixel changes from one class to another. We require the change to be persistent, meaning the new classification must be present for multiple time periods before generating a change detection. As a result change detections show change that occurred 3 weeks prior on average. Each detection includes a "score" percentage indicating how likely it is to be a true positive change so that users can determine their own tolerance for false positives.
Results
When filtering for detections with a score greater than 40%, our precision reached 0.75 for roads and 0.58 for buildings within Brazil, enabling users to more easily spot development and narrow the scope of where they needed to focus. We identified a few failure modes to address including agricultural areas, ships in port and coastal areas that surfaced significant numbers of false positives.
Conclusions
Planet has shown it is possible to automatically surface areas of human development over enormous areas with a high degree of precision, enabling users to more effectively use their resources to focus on areas that are changing.

Keywords:

Planet Machine Learning Computer Vision Roads Buildings Rainforest Automated

Justin Davis

Presentation within symposium:

S-23 Emerging uses of large-scale remote sensing in tropical forest monitoring

Detecting Infrastructure Growth Through High Cadence Satellite Imagery

-Review-