Canopy Height Estimation from Single Multispectral 2D Airborne Imagery Using Texture Analysis and Machine Learning in Structurally Rich Temperate Forests

Abstract

Canopy height is a fundamental biophysical and structural parameter, crucial for biodiversity monitoring, forest inventory and management, and a number of ecological and environmental studies and applications. It is a determinant for linking the classification of land cover to habitat categories towards building one-to-one relationships. Light detection and ranging (LiDAR) or 3D Stereoscopy are the commonly used and most accurate remote sensing approaches to measure canopy height. However, both require significant time and budget resources. This study proposes a cost-effective methodology for canopy height approximation using texture analysis on a single 2D image. An object-oriented approach is followed using land cover (LC) map as segmentation vector layer to delineate landscape objects. Global texture feature descriptors are calculated for each land cover object and used as variables in a number of classifiers, including single and ensemble trees, and support vector machines. The aim of the analysis is the discrimination among classes in a wide range of height values used for habitat mapping (from less than 5 cm to 40 m). For that task, different spatial resolutions are tested, representing a range from airborne to spaceborne quality ones, as well as their combinations, forming a multiresolution training set. Multiple dataset alternatives are formed based on the missing data handling, outlier removal, and data normalization techniques. The approach was applied using orthomosaics from DMC II airborne images, and evaluated against a reference LiDAR-derived canopy height model (CHM). Results reached overall object-based accuracies of 67% with the percentage of total area correctly classified exceeding 88%. Sentinel-2 simulation and multiresolution analysis (MRA) experiments achieved even higher accuracies of up to 85% and 91%, respectively, at reduced computational cost, showing potential in terms of transferability of the framework to large spatial scales.

A Learnable Model With Calibrated Uncertainty Quantification for Estimating Canopy Height From Spaceborne Sequential Imagery

“…Ground-truth CHM of the study area was acquired from LiDAR measurements with the Riegl 680i sensor in June 2017 and is depicted in figure 1. Details on the acquisition settings can be found in [11]. The ground sampling distance of the acquired 3D point cloud is 1m and the calculated CHM was bilinearly downsampled to 10m resolution.…”

“…As discussed in a following section, no cloud coverage filter is applied in any of the Sentinel-2 sequences. Finally, a land cover map of the area has been used for evaluation purposes and specifically in comparison with [10] and [11]. Based on the manual delineation of landscape patches (objects) [30], as created by local experts in 2012, we calculate aggregated pixel values of height in order to compare object-wise accuracy with the previous works.…”

“…Apart from pixel-wise evaluation, aggregations of pixels based on their vegetation type were classified in six vegetation height classes. In the study of Boutsoukis et al [11], SVM based classification of extracted features of texture was used to classify land objects in vegetation classes of the general habitat category taxonomy, based on high resolution airborne images on GSD smaller than 1m.…”

“…With this model architecture we are seeking to capture the temporal evolution of the distribution that the image texture follows and map it to the space of vegetation heights. Training and testing the model is based on different tiles of a total forest area of approximately 94.000 hectares and comparison results with state-of-the-art studies [11], [8], [10] are provided.…”

Urban forest canopy height modeling using spaceborne laser ICESAT-2 LIDAR

“…Therefore, the spatial distribution characteristics of urban elements and vegetation spatial environmental information should be considered when constructing urban tree height models. As a common measurement feature of the spatial distribution of urban elements, texture features are effective in inverting various tree parameters in optical remote sensing [8,9]. In addition, the current commonly used vegetation height calculation model mostly adopts a linear function [10,11], but the linear function is too simple, and the accuracy of the calculation results cannot be guaranteed.…”