Using AI to identify areas of deforestation and replanting

Client: Ministry for the Environment

Lynker Analytics with partners, Carbon Forest Services and UAV Mapping NZ Ltd were contracted by the Ministry for the Environment (the Ministry) in late-2019 to survey and classify around 7,500 distinct areas of potential forest loss across New Zealand.

cutover_ground.jpg

Deforestation is an important form of land-use change from a greenhouse gas perspective with assessments conducted in New Zealand every two years to meet international reporting obligations under the United Nations Framework Convention on Climate Change and the Kyoto Protocol.

In NZ, 40,000 – 50,000 hectares of forest is harvested each year with most areas replanted with a smaller proportion converted to another land use.

Between January and August 2020, the Lynker Analytics Consortium conducted an aerial survey of New Zealand using Cessna 172 aircraft flying at approximately 5,000 feet above ground level capturing over 30,000 high resolution vertical photographs with a spatial resolution of 0.25m.

forestrysurvey4.jpg

The imagery was georeferenced and then classified into land cover classes such as cutover (harvest area), plantation seedlings, grass/pasture, mature exotic and mature native forest using Machine Learning (ML). 

Multi-scale convolutional neural network (CNN) models were then used to classify each target into land cover units with a spatial resolution of 100 square metres. In deep learning, a CNN is a class of deep neural network that is most commonly applied to analysing visual imagery.

Over 15,000 multi-scale training annotations were used to train the models. The example below shows a training example with the left-hand image showing a 70m x 70m area of new plantation forest with the red box representing the right-hand image, which is a smaller area of 10m x 10m. 

seedlings1.png

For the neural network models, a patch-segmentation model was trained on the multi-scale image chip pairs.  Patch segmentation was used instead of U-Net (or similar) for semantic segmentation as this allowed us to gather training annotations quickly and it better described land use in forested areas by revealing planting patterns with individual trees separated by scrub or cutover.

From this we applied a geospatial data generalisation routine to filter out noise or speckle and then we used a multi-criteria iterative analysis to assign each area into dominant land cover categories broadly in line with the Emissions Trading Scheme - Geospatial Mapping Information Standard. 

Overall we achieved a combined overall model accuracy approaching 80% which given the heterogeneity in land cover is a good result which then enabled the model to be used as a land cover change detection system in concert with the multi-criteria analysis. Examples of the final classified results are shown below.

results1.png
result4.png
result3.png
result2.png

The automated monitoring system proved reliable in detecting deforestation, re-planting and other land cover changes exceeding one hectare.  It also enabled more rapid assessment of replant status used by the Ministry for reporting.   In total 7473 targets were surveyed with imagery, land cover polygons, dominant land cover and replant status attributes provided to the Ministry. 

Watch the officlal presentation of this report to MfE here. The technical report is available here.

Share