Mapping the global potential of natural reforestation projects | by Steve Klosterman | Aug, 2023

using ground observations, remote sensing, and machine learning

By Stephen Klosterman and the Earthshot Science Team. Content originally presented at the American Geophysical Union Fall Meeting, in December 2022.

Ecological restoration projects often require investment to get activities up and running. In order to create carbon finance opportunities for forest growth and conservation projects, it’s necessary to be able to predict the accumulation, or avoided emission in the case of prevented deforestation, of carbon in woody biomass. This is in addition to trying to understand the likely changes in a wide array of other ecosystem properties, e.g. plant and animal species composition and water quality. In order to create carbon accumulation predictions, a common approach is to devote individual attention and research effort to projects in specific locations, which may be scattered across the globe. It would therefore be convenient to have a locally accurate and global map of growth rates, or other parameter values of interest, for fast “prospecting” work of determining ecosystem restoration opportunities. Here we describe methods to create such a map, derived from a machine learning model trained on data from a previously published literature review. We then demonstrate the implementation of the map for Africa in a Google Earth Engine app.

We used a recently published dataset of forest stand biomass measurements, ages, and geographic locations (Cook-Patton et al. 2020) to train a machine learning model to predict a parameter of the commonly used Chapman-Richards (CR) growth function.

After cleaning the data of outliers and unrealistic observations similar to what was done in the original publication, we were left with about 2000 observations, shown here on a global map with symbol size proportional to the number of observations per site:

Global distribution of site-based data; symbol size proportional to number of measurements per site. Image by the author.

‍The observations were spread across 390 sites. Most sites (64%) just have one measurement, while there is one site that has 274 measurements.

Source link

Leave a Comment