Mangrove VM33, VM7 Explainer: How do you detect mangrove extent using machine learning techniques?

Within mangrove projects (and any project), it goes without saying that accurate baselining is absolutely key. In other words for conservation projects, how many mangroves have we actually lost over the last 10 years? Or, how have mangroves in the area been naturally restored for reforestation projects?

One of the best ways to ensure baselining is done well is to build a credible localised mangrove detection model with remote sensing.

To start, take a look at Image (a) below. Based on the dark green land mass pictured, how do we know which trees are mangroves, and which ones are not mangroves?

Figure 1. Image from https://www.mdpi.com/1999-4907/15/1/127

Now, machine learning models aren’t perfect. Upon closer inspection, you can see the roads are not drawn correctly by the model, in figure C1 and C2. But, this is a powerful illustration to demonstrate how a machine can tell you very likely where mangroves are for a particular year with a 80 - 90% accuracy.

What is remote sensing?

Remote sensing are satellite-based technologies that launch invisible sensors into the earth’s surface. The manner in which the sensors interact with the ground (such as how certain sensors are sensitive to the chlorophyll in trees) or distance it takes to bounce off (the height) can tell us a lot about what’s on the ground.

Figure 2: What is remote sensing https://crisp.nus.edu.sg/~research/tutorial/optical.htm

If we know which types of sensors are sensitive to mangroves, and are then able to train a machine learning model (specifically in this case, a supervised machine learning model) to draw what are mangroves from all the data these sensors produce. A supervised machine learning model is a type of model that is trained on labelled datasets.

First, we need to try and get a dataset for at least some known dataset from native mangrove species that we are confident in. This usually includes downloading local university or governmental datasets with land use and classification data for a particular year to create datasets for where a particular mangrove species are. As always with machine learning and data, the more good data (peer reviewed) the better.

Figure 3: Example of a peer review land use classification data: https://www.mdpi.com/2072-4292/16/1/76

This is what a good dataset may look like. You can see this is likely hand labelled, which adds greater confirmation as a highly accurate dataset to train a model on.

From these types of datasets, you need to create two very important datasets.

The first is called your “positive” training dataset. In other words, what you know to be very likely the mangrove species you are looking for. This will entail segmenting out the areas such as “Closed Mangrove” and “Open Mangrove” above as a “positive” training dataset.

The second is called “negative” datasets. Similar to how humans learn and mature, it is equally important to know what mistakes not to make for a machine. In other words, what you know to very likely “not” be the mangrove you are looking for. 

Below is an example that demonstrates why it is very important to include areas that may commonly be mistaken by machines to be mangroves areas.

Figure 4. Left: Actual satellite image. Right: Inaccurate machine learning detection of 'mangroves' in orange

Upon closer inspection on the results above, we can see where areas of the machine learning model are incorrectly detecting the shrimp farms as mangrove.

This is a great example of where there may be a lack of sufficient “negative” datasets for what aren’t mangroves. One way we may further fine tune or improve the model is by adding more of such datasets of shrimp ponds, paddy fields and non-mangrove forests from Figure 4 above as a “negative” training datasets, which will likely improve the mangrove detection model further. Of course, if one has the resources, adding in ground truth data is always the gold standard.

Once the model has acceptable results, one can re-deploy this model across different datasets in different years for the same area to determine what the mangrove extent was like for that year. Giving a developer or financier, a powerful indicator for what the baseline is truly like. While there are certainly other methods, what we have shown you today is one of the most deployed methods by experts for mangrove detection for feasibility studies when doing mangrove projects.

Looking to assess your mangrove or peatland project? Nika.eco are carbon modelling specialists. It is our core business to deeply understand carbon methodologies using advanced machine learning to automate highly complex modelling calculations.

Reach out if you require carbon due diligence services or looking for a Technical Partner on your projects.