Habitat Classification of Terra Preta Forest in Amazonia using Google Earth Engine. Eventually modified and published in Nature Sustainability, 2024
The goal of the Terra Preta project is to better estimate the distribution and abundance of Terra Preta soils. These human-modified soils, or anthrosols, are dark due to their charcoal content along with other biomarkers. Whether used as a refuse middden, a slash-and-burned field, or a yard for playing, any Amazonian soils that exhibit characteristic traits of Terra Preta are deemed as such. With better estimates of Terra Preta comes refined carbon estimates for the Amazon rainforest, a potential progression of this project.
To identify areas of potential Terra Preta, code was executed in Google Earth Engine where real-time, interactive displays of outputs on maps makes for a convenient iterative framework. User-supervised classification techniques, whereby areas of known habitat classes are used to train a model on how to identify other similar classes. A random forest machine learning method proved to be a useful local classifier, with high resubstitution accuracy for trained areas, however its accuracy in predicting Terra preta external to the training areas has yet to be definitively tested.
Due to extensive fire scarring in the region, mainly resulting from prescribed burns, the model originally confused the spectral signature of this disturbed area with Terra Preta. To limit the impact of fire scars, a difference in the Normalised Burn Ratio Thermal (dNBR) was used mask out areas of imagery with large differences in pre and post-fire burn indices. Additionally, because Terra Preta establishments are rarely found within the active floodplain of the Xingu River, we masked out the entire floodplain to reduce training noise. By adjusting the tolerance for dNBR in the "burn-removal" script, the influence of burned areas may be further reduced. In masking out areas of forest, no classification can be made for that area. By assigning the mode value of surrounding cells to masked cells post-classification, a continuous version could be attained.
Other potential ways the prediction of Terra Preta could improve include running various models and classifiers and seeing where they most often agree. Also, further research into the precise spectral signatures produced by Terra Preta, along with their seasonal dependence, may help refine estimates. McMichael et al., 2014 will provide good insights to this end.
Correspondance can be made with me. Work on the project was completed in mid-early 2020 with the Perron Geomorphology group at Massachussets Institute of Technology (MIT) and is available under an MIT license with the exclusion of the training data.
Scripts included: Classification: used to train a classifier with Landsat imagery, then run the classifier on the imagery and assign classvalues to pixels. burn-removal: used to produce a clipping layer that can remove burn scars from imagery in pre-classification