Marketplace

Tree Detection

Finds trees in the image and for each pixel returns the probability map of it to be a part of a tree.

Description

Tree Detection identifies all pixels corresponding to trees in satellite images. The block runs on Pleiades Streaming (recommended) or SPOT Streaming Data Blocks and outputs a shadow mask with the probability of each pixel corresponding to a tree. NB: We recommend using 0.51 for this particular model following in-field tests for powerline inspection applications.

The solution uses deep learning techniques for computer vision and is constantly improved with subsequent updates; thus far, it has been trained on tens of thousands of semantically annotated images. The model is used as part of the Tree Height Detection from Shadows block.

Due to its complexity, this analysis block requires a significant running time. There are no restrictions on image dimensions, but the computing time will increase with the size of the image. We recommend using the Raster Tiling block upstream to speed up processing and reduce infrastructure costs for large areas.

Use cases include Vegetation Management, Urban Planning & Construction, Fire Risk Estimation, Land Use and Management, Forestry, Powerline inspection, Trainline inspection.

Note: This block currently can be used with Pléiades or SPOT Streaming followed by the Raster Tiling Block. Future data block integrations will come in the next versions. Contact UP42 if you have suggestions for which data blocks should be prioritized first.

General InformationDescription
Block TypeProcessing for Shadow Detection
Supported input dataRaster Tiled (512x512) GeoTIFF images.
Output data formatGeoTIFF heat map image containing the probability for each pixel to be part of a tree.
ResolutionSame resolution as the input data.
PerformanceThe model accuracy ranges from 80 to 90% depending on input block resolution and location. This is real-world accuracy following benchmarking against helicopter based LIDAR. Benchmarked against human annotators on 10,000+ Pleiades images the model showed an average precision of 91%, recall of 94.5% and an F1 score of 93%. Please note that for seasonal trees, the block works best when the input images have the trees in full foliage. The model aims to focus on medium to large trees, limiting detection confidence for brushes and ignoring high grass.

Video tutorial

We have created a 5-minute demonstration on running an analysis process using the Spacept Tree Detector Block, which is accessible on YouTube here.

Example Input

{
   "oneatlas-pleiades-aoiclipped:1":{
      "ids":null,
      "bbox":[
         -1.039205,
         46.732618,
         -1.026172,
         46.743978
      ],
      "time":"2018-01-01T00:00:00+00:00/2020-12-31T23:59:59+00:00",
      "limit":1,
      "zoom_level":18,
      "time_series":null,
      "max_cloud_cover":100,
      "panchromatic_band":false
   },
   
    "tiling:1":{
      "nodata":null,
      "tile_width":512,
      "tile_height":512,
      "match_extents":false,
      "output_prefix":"",
      "augmentation_factor":1,
      "discard_empty_tiles":true
   },
   
    "tree-detection:1":{
   }
}

Terms & Conditions

View the End User License Agreement conditions.

Ready to get started?

Get in touch or become a partner.

Contact salesBecome a partner