Mapping tree crop plantations in the Amazon using probabilistic satellite image-based reference data collection

Keywords: Brazil, Tree crops, Satellite, Strata, Sampling, Interpretation
Abstract Type: Poster Abstract

Abstract

The goal of this project is to support national efforts to monitor agricultural expansion in the Amazon by providing new information on the distribution of tree crop plantations within existing agricultural lands. To achieve this, we developed a 1) probabilistic, satellite image-based reference data collection dataset that distinguishes key tree crops (e.g. oil palm, cacao) from other land uses in cultivated regions, and 2) used this to train and assess tree crop mapping models.

First, point samples were generated using a Google Colab notebook connected to the Google Earth Engine to stratified sampling approach designed to optimally allocate samples when the class of interest is rare (<10% of the study region). The strata were based on Terra Cover, an existing 10 m land cover map for the Para State of Brazil, which was used to define two primary strata, one a superclass created by combining classes likely to represent the tree crop(s) of interest (e.g. silviculture, woody pasture, perennial crops, and other semi-perennial crop classes), and the other comprised of classes representing tree crop absence (e.g. secondary regrowth, non-crop classes, and annual cropping systems). Following selection, the sampled points were loaded into Collect Earth Online to create a project that was conduct image interpretation-based data labels, using a multi-labeller approach that estimated label uncertainty. These points were divided into training, validation, and testing subsets to develop an R-based Bayesian model that uses the SAR-optical image stack to predict the probability of tree crop presence throughout the agricultural matrix.