Can 2019 GEDI data be used to tune catastrophic deforestation model?

[crich011]

Given a height measurement and other canopy data from 2019 GEDI observations and observed 2018 or 2019 deforestation events from Landsat (verify appropriate time overlap), can we build a model to identify events?

It is possible that this is beyond the scope of MVP work, but I think having the full waveform GEDI data to ground truth our analysis and the Landtrendr outputs (plus any other features we want to include) could allow us to build a really good model.

Some considerations relating to the data:

1. We need to ensure that we are using GEDI observations that occur after any candidate events; depending on the time window associated with GEDI data, that might mean using earlier (e.g. 2018) deforestation events as our set of interest.
2. We would need to set thresholds for identifying regions to include and somehow randomize over these regions of interest.

Some considerations relating to the model/analysis:

1. Currently, we use ltgee.getChangeMap to identify deforested regions. This takes the output of the Landtrendr algorithm as an input and applies some filtering subject to user-provided hyperparameters. Ideally, we would train a model using the quantities getChangeMap cuts on as input features.
2. Unfortunately, I have yet to have any luck finding the source code for getChangeMap, but it might be possible to reproduce something like it (or close enough to get inspiration) by looking at the Landtrendr outputs.
3. It's also very possible that someone has already done this, so a healthy amount of Google searching is a good idea.
4. Model simplicity and interpretability are important considerations here, namely because a simple model that uses only Landtrendr outputs and can provide (potentially regionally-varying) best fit values for getChangeMap hyperparameters as an output means we can use getChangeMap directly, as opposed to having to apply a CNN or XGBoost model ourselves.

[crich011]

With regard to the points about getChangeMap source code, while I still have yet to find the actual implementation, the tutorial here all but walks users through creating a getChangeMap function. The one piece missing is the calculation of the disturbance signal-to-noise ratio (DSNR), and I am unable to access the full text of the paper, but based on what I can find, it appears to just be computed as the magnitude of the change divided by the RMSE for that band.

[crich011]

I was able to use getSegmentData to generate the segments that feed getChangeMap, including DSNR. These images are array-valued, and using array manipulation and ee.Image.arrayMask, I was able to build a function which masks out all but single-year events for an arbitrary year. Masking out entries in the array-valued image, before converting to bands, is more performant because it means we have to convert fewer entries bands.

Quantities which we need to fix in order to define our input features (e.g. year, duration, initial index magnitude) can be filtered before converting the arrays to bands. Quantities which we wish to tune or determine using the model (e.g. magnitude of the decrease) should be sent to bands. Because of the way we compute it, I am unsure of the best way to handle something like minimum patch size.

The two somewhat sensible patch size approaches I can think of are:

1. Construct an ensemble of N training sets, where each set has a different patch size filter applied before training; this would essentially be filtering on the existence of a decrease, but not making cuts on its magnitude.
2. Train a model irrespective of patch size (I am unsure of how, if at all, standard methods in Earth Engine consider neighboring pixels), and apply post hoc filtering based on patch size to determine our desired value.

[crich011]

This work is currently blocked on the work of #8/#9, but once #9 merges and #2 finishes, we will be able to work on implementing a version of the deforestation model on the full GEDI data.