Signature extension as a machine learning strategy for mapping plantation forest genera with Sentinel-2 imagery

Abstract

Plantation forest inventory data such as tree genus is important for supporting forestry decisions and policymaking, but such data are traditionally collected in-field, which is time-consuming and costly. Although machine learning and remote sensing technologies have shown great potential to reduce the time and effort for mapping forest plantation genera at regional scales, they rely on training (labelled) data, which are also costly to collect over large areas. One approach to reducing the effort of training data collection is to make use of signature extension, whereby training data collected in one area is used to train and apply a machine learning model in a different area. This study aimed to evaluate the viability of training data signature extension for constructing random forest (RF) machine learning models to differentiate between acacia, eucalyptus and pine trees using Sentinel-2 imagery as input. The study was carried out over a large area (about 4920 km2) in South Africa. The study area was divided into 19 tiles of 100 × 100 km (each tile coincides with the footprint of a Sentinel-2 tile) from which three were chosen for sourcing (collecting) training data. Four experiments were conducted. In the first experiment, a fixed number (3000 per genera) of training samples were collected in the first source tile and used to build an RF model. The resulting model was then applied and assessed in all 19 tiles. This protocol was repeated in the second and third experiments using the training data collected in the other two source tiles respectively. In the final experiment, training data from all three source tiles were combined and applied to all 19 tiles. The mean overall classification accuracy of each classified tile was compared to the extension distance (i.e. the distance between the target and source tile), differences in rainfall seasonality, and variation in the mean annual temperature among tiles to gain an understanding of how signature extension efficiency is influenced by distance and environmental conditions. The results show that signature extension is viable (∼70% overall accuracies) over distances of up to 500 km, but only if the source and target tiles represent areas with similar rainfall regimes.