forest_monitoring

Forest Monitoring - Road Development

Note: These notebooks currently use deprecated versions of our Python SDK, as well as PSOrthoTile, an item-type that is no longer available. An updated version of these notebooks are currently in progress.

In a region in the Democratic Republic of Congo, road development caused loss of forests between September and November 2017. In this notebook project, we demonstrate an end-to-end use case where we have an Area Of Interest (AOI) and want to detect the deforestation due to road development.

Because this is a large use case, it is covered in multiple notebooks. Each notebook is made to work stand-alone but some start with information obtained and saved in previous notebooks.

The primary workflow for this use case utilizes PSOrthoTiles and is implemented in the following notebooks, in order:

1. drc_roads_download: Find and prepare images that overlap AOI
2. drc_roads_classification: Classify an image of the AOI into forest and non-forest regions
3. drc_roads_temporal_analysis: Identify change in the AOI (new roads being built) using temporal analysis.

The workflow above uses the Unusable Data Mask (UDM) to determine scene quality and filter out bad pixels. Since that workflow was implemented, the Usable Data Mask (UDM2) was released. This asset provides additional information on pixel usability. To assess the impact of UDM2 on forest/non-forest classification, the entire workflow has been implemented utilizing UDM2 in the following notebook:

• drc_roads_udm2

Because UDM2 was just recently released, the notebook had to restrict it's search to recent imagery, resulting in less useful imagery being found.

Additionally, the entire workflow has been implemented utilizing mosaics and is implemented in the following notebook:

• drc_roads_mosaic

Change classification results for the PLOrthoTiles and mosaic data inputs are very promising. With some clean up, we could have a pretty cut and dry change classification. Classification of forest/non-forest regions was better with mosaics than PSOrthoTiles. This is likely due to haze in the PSOrthoTiles that isn't identified in the UDM. Unfortunately, UDM2 did not pick up the haze in the few available images with UDM2, so classification quality was not improved. Mosaic pre-processing removes the inconsistencies due to haze and other factors, which did improve classification quality.

Techniques Demonstrated

There are many different techniques demonstrated in these notebooks:

1. drc_roads_download:
   • Identification of PSOrthoTile strips that significantly overlap the AOI
   • Activation and download of PSOrthoTiles as cloud-optimized geotiffs (COGs), mosaicing into single strip images
   • Use of planet client downloader to activate, download, and mosaic multiple scenes across multiple strips
2. drc_roads_classification:
   • Classification of an orthotile strip image into forest and non-forest regions using unsupervised (KMeans) and supervised (Random Forests) techniques
3. drc_roads_temporal_analysis:
   • Classification of pixels within AOI as change or no change based on temporal analysis of the forest/non-forest classified imagery
4. drc_roads_mosaic:
   • Re-use of code from other notebooks
   • Activation and download of mosaics using GDAL with the PLMosaic driver
   • Georeferencing and projecting, resampling, and cropping a label image to match the mosaic
   • Creating a training dataset from a mosaic and a label image
   • Classification of mosaic pixels into forest/non-forest using Random Forests classifier
   • Classification of forest/non-forest images into change/no-change using Random Forests classifier