Testing various vegetation indices based on RGB+NIR bands
The statistical results were compared to the results of image classification. Maximum Likelihood Classification (MLC) gives the best results compared to other possible classifiers (e.g. non-parametric classifiers such as random forest, support vector machine, decision trees etc.). Samples of healthy and infested trees were selected randomly in total counts of 10:40 for training as well as for validation. Classification was processed in ArcGIS using the following inputs: Greenness Index (GI), Simple Ration (SR), Green Ratio Vegetation Index (GRVI), Normalized Difference Vegetation Index (NDVI) and Green Normalized Difference Vegetation Index (GNDVI).
The healthy (H-sample) and the infested (G-sample) trees were compared by statistical analysis. Infested trees were included in the G-sample, while H-sample was a random sample of the same size selected from the set of healthy. For the graphical comparison of the selected samples, see figures below.
Box-plots for SR, GI and GRVI, respectively (always in the order G-sample - infested, H-sample - healthy):
Box-plots for NDVI and GNDVI, respectively (always in the order G-sample - infested, H-sample - healthy):
Spectral curves of healthy, infested and dead trees were described
A complete spectral curve of the infested trees over time was extracted from the imagery, i.e., a description of spectral changes in the characteristics of the infested trees from the healthy state up to its death. Results obviate that such changes can be observed even by low-cost means soon after the tree infestation. A tailor-suited semiautomatic algorithm developed within the project achieved approx. 80% accuracy in early detection of infested trees and the algorithm was almost infallible (99% accuracy) in detection of dead trees infested in previous seasons/bark beetle generations.
Processing workflow estabilished
In this project, regular UAV imaging of the forest stand in the study area was performed to investigate a possibility of early detection of changes in the spectral characteristics of infested trees using a low cost solution (standard digital camera and UAV) before obvious visual changes occur. The detection is therefore based on the study of changes of spectral characteristics of the infested trees over time. Subsequently, the feasibility of acquired low-cost solution was verified through detection of infested trees using advanced image analysis (utilizing both standard remote sensing and photogrammetry techniques).
Ongoing statistical analysis - spectral reflectance of infected trees
Look at the running statistical analysis of our bark beetle project. See below boxplots that represent a progress of spectral characteristics of Red and Near-IR bands of infected trees across the season.
Masking of shaded parts of trees
The next step of image processing, which was necessary before detection of trees attacked by bark beetle, was masking shaded parts of tree crowns and trees lower than 5 m (an assumption that bark beetle infested spruce trees older than 60 years). The selected workflow was based on finding an ideal threshold for distinguishing shadows in the near-infrared band and their combination with a binary mask created from the normalized Digital Surface Model (nDSM) in ENVI (version 5.5) software.
Treetops and crowns delineation
The crucial part of detection of trees attacked by bark beetle, which was solved in the past days, is treetops (red points) and crowns (yellow objects) delineation. The solution using the normalized digital surface model created from UAV images by Structure From Motion algorithm and methodology based on invert watershed approach was chosen.
