Abstract
The effective identification of ground objects from aerial platforms is a significant challenge in forested environments, where dense canopies can impact sensor performance. While the negative impact of vegetation on remote sensing is well-documented, the relationship between specific canopy structures and sensor performance degradation has not been quantified. This study addresses this gap by assessing how vegetation path length affects the accuracy of an object classification model. To achieve this, high-resolution aerial imagery and lidar data were collected over the Harvard Forest, where 68 ground targets were placed under varying levels of canopy occlusion. Lidar-derived point clouds were used to model the forest structure and calculate the sensor-to-target path length through different vertical layers of the canopy. A Convolutional Neural Network (CNN) was then trained to identify the cardinal orientation of these targets from aerial images, with its classification accuracy serving as a proxy for object identification performance. The results demonstrate a statistically significant negative correlation between classification accuracy and vegetation path length. A piecewise regression analysis identified a critical threshold at a canopy path length of approximately 10 meters and a total path length of 25 meters, beyond which classification accuracy diminished significantly. These trends were consistent across deciduous, coniferous, and mixed forest types. This study's quantitative framework helps assess sensor performance in complex vegetated environments, informing applications such as environmental monitoring, search and rescue, and autonomous navigation.

This publication has 19 references indexed in Scilit: