Predicting Eucalyptus stand attributes in Minas Gerais State, Brazil: an approach using machine learning algorithms with multisource datasets

Abstract

Quantitative spatial information on forest attributes is critical in forest management as an important indicator of biophysical processes, forest dynamics and the provision of services and goods. In this thesis, the effectiveness of integrating field data, multispectral optical imagery obtained from Landsat 5 Thematic Mapper (TM), Landsat 8 Operational Land Imager (OLI), and Sentinel-2A satellites, synthetic aperture radar (SAR) data acquired by the Sentinel-1B satellite, digital terrain attributes derived from a digital elevation model (DEM) and climate data was tested using parametric and nonparametric methods of spatial prediction for estimating and mapping forest stand attributes in Eucalyptus plantations in northern Minas Gerais state, Brazil. For this purpose, this thesis was organized in four articles. In the first one (1), Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Networks (ANN) methods were assessed to estimate and map basal area and volume using Landsat 5 TM data. RF showed the best performance for spatial prediction and mapping of stand attributes in Eucalyptus stands, and for this reason, it was used in the next three articles. In the second article (2), different combinations of stand age with variables extracted from three different digital datasets (i.e., Landsat 8 OLI multispectral optical data, Sentinel-1B SAR data, and DEM-derived terrain attributes) were tested to estimate volume. The results showed that the best data combination corresponds to the integration of all datasets (i.e., stand age and the selected variables of Landsat 8 OLI and Sentinel-1B SAR imagery, and DEM-derived terrain attributes). The third article (3) investigated the potential of Sentinel-2A multispectral information for improving forest attribute estimates compared with Landsat-8 OLI imagery when both multispectral optical imagery (i.e., Sentinel-2A and Landsat 8 OLI) were combined with Sentinel-1B SAR data and DEMderived terrain attributes. As expected, the Sentinel-2A optical data appeared to have a greater explanatory power in predicting forest attributes of Eucalyptus plantations compared to Landsat 8 OLI imagery. In the fourth article (4), a nonparametric modeling approach was used to examine relationships between terrain attributes and climate data on forest site productivity and maximum mean annual increment (MAI max ) of Eucalyptus plantations at a regional scale. Terrain attributes and bioclimatic variables showed good potential to classify site productivity and to predict MAI max in our study area. The large and increasing area of Eucalyptus forest plantations in Brazil and elsewhere suggest that the new approaches developed here to estimate forest stand attributes and productivity have great potential to support Eucalyptus plantation monitoring and forest management practices.