Browsing by Subject "Forest height estimation"

Accurate forest height estimates are critical for environmental, ecological, and economical reasons. They are a crucial parameter for developing forest management responses to climate change and for sustainable forest management practices, and are a good covariate for estimating biomass, volume, and biodiversity, among others. With the increased availability of Light Detection and Ranging (LiDAR) data and high-resolution images (both satellite and aerial), it has become more common to estimate forest heights from the sensory fusion of these instruments. However, comparing recent advancements in height estimation methods is challenging due to the lack of a framework that considers the impact of varying data resolutions (which can range from 1 meter to 100 meters) used with techniques like convolutional neural networks (CNNs). In this work, we address this gap and explore how resolution affects error metrics in forest height estimations. We implement and replicate three state-of-the-art convolutional neural networks, and analyse how their error metrics change as a dependency of the input and target resolution. Our findings suggest that as resolution decreases, the error metrics appear to improve. We hypothesize that this improvement does not reflect a true increase in accuracy, but rather a fundamental shift in what the model is learning at lower resolutions. We identify a possible change point between 3 meter and 5 meter resolution, where estimating forest height potentially transitions to estimating overall forest structure.