Construction of 3D maps of vegetation indices retrieved from UAV multispectral imagery in forested areas

The construction of fuel moisture content (FMC) maps, as well as temperature, terrain topography, and wind speed maps, are essential for the development of fire susceptibility models in forested areas. Moisture distribution in tree canopies requires exploration and a three-dimensional representation. This paper presents the construction of FMC maps expressed as vegetation indices (VIs) in a point cloud. Multispectral images were captured by a camera mounted on an unmanned aerial vehicle to create the point cloud. VIs were estimated in the points that belonged to the forest canopy. To classify the canopy points, we a combination of filtering of ground points and thresholding of VIs was evaluated. On such canopy points, random forest (RF), kernel ridge regression (KRR), and Gaussian pro-cess retrieval (GPR) regressors were investigated to estimate twelve VIs related to FMC. The input set of the models consisted of the points representing five wavelengths provided by the multispectral camera. The ground truth of VIs was obtained using a spectrometer. The study area was a 1 ha forest of Pinus radiata in the Maule Region, Chile. The results demonstrated that combining ground filtering and VIs thresholding for canopy points segmentation achieved a precision of 93.27%, recall of 95.65%, F1 score of 90.12%, and ac-curacy of 87.82%. Furthermore, the recovery of the VIs using GPR achieved a root mean square error of 0.175 and a coefficient of determination of 0.18. According to the correlation coefficient, GPR was able to recover eleven of the twelve VIs, KRR recovered three, and RF failed to recover any.(c) 2021 IAgrE. Published by Elsevier Ltd. All rights reserved.

Automated summary

This paper presents the construction of fuel moisture content (FMC) maps expressed as vegetation indices (VIs) in a point cloud for the development of fire susceptibility models in forested areas. Multispectral images captured by a camera mounted on an unmanned aerial vehicle were used to create the point cloud, and VIs were estimated in the forest canopy points. The results showed that the combination of ground filtering and VIs thresholding achieved high accuracy and recall rates for canopy points segmentation. Gaussian process retrieval (GPR) showed good performance in recovering the VIs.