Seeing the forest for the trees

At Abley we’ve been working with the Ministry for Primary Industries to map exotic forest from satellite imagery.

Extracting information with satellite imagery

Satellite imagery provides an objective and timely way for organisations to extract information about features on the earth’s surface. A fascinating example of this is the non-profit Global Forest Watch who provide near real-time monitoring of global forest cover. Deforestation alerts have allowed officials to take rapid action against illegal logging in the Amazon.

In New Zealand, the National Exotic Forest Description (NEFD) is an annual survey of forest owners and managers. The survey informs an authoritative, high-quality record of the country’s production forests. By extracting exotic forest land cover from satellite imagery, the survey results can be tested against an independent dataset and kept up to date.

Classifying exotic forest

We completed a feasibility study to explore options for creating a high spatial and temporal resolution exotic forest dataset. Using Esri’s machine learning and deep learning toolsets we were able to quickly prototype techniques for classifying exotic forest in the Canterbury Region. These tools provide an end-to-end deep learning workflow within ArcGIS Pro. The team ran the process with ArcGIS Notebooks (Esri Python environment) to build a repeatable process, but they are also available as standard geoprocessing tools.

Our assessment trialled both the Support Vector Machine classifier and U-net architecture with Sentinel-2 satellite imagery. Multispectral imagery measures light in a number of spectral bands (Sentinel-2 has 13 bands). Materials (like buildings, trees, lakes) reflect and absorb light differently and so by defining their ‘signature’ against multiple bands we can more accurately define and classify land cover.

A key challenge for us was differentiating the spectral signatures of exotic forest species from indigenous species. This is due to the spectral similarity of these features. The classification workflows required creation of training data, training of models, testing, and analysing accuracy to ensure spectral variability was captured. Multiple iterations were completed to improve the overall classification accuracy.

It was determined that the U-net performed more accurately, although it did fail to classify many small areas of forest, whereas Support Vector Machines over-classified (in places confusing native forest and arable cropland for exotic forest). With more time, additional training data and model training would produce more refined results.

Demonstrating success

Utilising these powerful Esri toolsets we were able to quickly demonstrate success in creating a high spatial and temporal resolution exotic forest dataset. Ultimately, the client was pleased with the results and could see value in this approach to streamlining the acquisition of information about exotic forest in New Zealand.

Contact Andy to learn more about this project and the Esri machine learning and deep learning toolsets.

exotic forest image
U-net exotic forest classification