Species-specific machine learning models for UAV-based forest health monitoring: Revealing the importance of the BNDVI

Abstract

Exploring the capabilities of remote sensing technologies for identifying stress responses in trees due to environmental pressures is crucial for comprehension, management, and maintenance of forests that are productive, healthy, and resilient. In recent decades, research on forest health monitoring has been predominantly focused on data obtained remotely, either from satellites or crewed aircraft. During the last few years, Uncrewed Aerial Vehicles (UAVs) have gained prominence as invaluable remote sensing platforms, increasingly being employed for forest surveying. As intermediary between traditional remote sensing methods and ground-level observations, UAVs can capture high-resolution imagery from low altitudes, even below cloud cover, in unprecedented detail. This ability allows for the precise detection of stress responses at the individual tree scale. In our study, we have acquired a highly heterogenous, multispectral time-series dataset from the International Co-operative Programme on Assessment and Monitoring of Air Pollution Effects on Forests (ICP Forests) inventory plots across Bavaria, Germany, focusing on the main tree species. The data was recorded over three consecutive years from a UAV with the objective of monitoring tree physiological stress responses. Concurrently, with the drone flights, the ground-based forest condition surveying (Level-1 monitoring) was conducted, serving as ground-truth validation, and providing detailed information on tree health indicators, such as defoliation and discoloration. Our findings revealed that multispectral imagery obtained from a UAV closely aligns with field data, proving effective detection of physiological stress in trees. Remarkably, in conjunction to the red, red edge, and near-infrared band, the inclusion of the blue band emerged as a critical indicator of tree stress when incorporated into the Blue Normalized Difference Vegetation Index (BNDVI), depending on factors such as tree species, class division, and atmospheric conditions. Furthermore, the averaged values per sample tree over three years, alongside the 5th and 25th percentile of the data distribution, have proven to be of key importance. Based on spectral indices, we achieved good classification accuracies by training species-specific gradient boosting models (macro F1-scores ranging from 0.492 to 0.769). These models can assist in quantifying tree stress responses, thereby supporting the objectives of the ICP Forests program, potentially leading to substantial cost savings or increased coverage in the future.