Detecting overmature forests with airborne laser scanning (ALS)

Building a network of interconnected overmature forests is crucial for the conservation of biodiversity. Indeed, a multitude of plant and animal species depend on forest structural maturity attributes such as very large living trees and deadwood. LiDAR technology has proved to be powerful when assessing forest structural parameters, and it may be a promising way to identify existing overmature forest patches over large areas. We first built an index (IMAT) combining several forest structural maturity attributes in order to characterize the structural maturity of 660 field plots in the French northern Pre-Alps. We then selected or developed LiDAR metrics and applied them in a random forest model designed to predict the IMAT. Model performance was evaluated with the root mean square error of prediction obtained from a bootstrap cross-validation and a Spearman correlation coefficient calculated between observed and predicted IMAT. Predictors were ranked by importance based on the average increase in the squared out-of-bag error when the variable was randomly permuted. Despite a non-negligible RMSEP (0.85 for calibration and validation data combined and 1.26 for validation data alone), we obtained a high correlation (0.89) between the observed and predicted IMAT values, indicating an accurate ranking of the field plots. LiDAR metrics for height (maximum height and height heterogeneity) were among the most important metrics for predicting forest maturity, together with elevation, slope and, to a lesser extent, with metrics describing the distribution of echoes' intensities. Our framework makes it possible to reconstruct a forest maturity gradient and isolate maturity hot spots. Nevertheless, our approach could be considerably strengthened by taking into consideration site fertility, collecting other maturity attributes in the field or developing adapted LiDAR metrics. Including additional spectral or textural metrics from optical imagery might also improve the predictive capacity of the model.