Why can we detect lianas from space?

Abstract

Lianas are found in virtually all tropical forests and have strong impacts on the forest carbon cycle by slowing tree growth, increasing tree mortality and arresting forest succession. In a few local studies, ecologists have successfully differentiated lianas from trees using various remote sensing platforms including satellite images. This demonstrates a potential to use remote sensing to investigate liana dynamics at spatio-temporal scales beyond what is currently possible with ground-based inventory censuses. However, why do liana-infested tree crowns and forest stands display distinct spectral signals? And is the spectral signal of lianas only locally unique or consistent across continental and global scales? Unfortunately, we are not yet able to answer these questions, and without such an understanding the limitations and caveats of large-scale application of automated classifiers cannot be understood. Here, we tackle the questions of why we can detect lianas from airborne and spaceborne remote sensing platforms. We identify whether a distinct spectral distribution exists for lianas, when compared to their tree hosts, at the leaf, canopy and stand scales in the solar spectrum (400 to 2500 nm). To do so, we compiled databases of (i) leaf reflectance spectra for over 4771 individual leaves of 539 species, (ii) fine-scale (∼1m²) surface reflectance from 999 tree canopies characterized by different levels of liana infestation in Panama and Malaysia, and (iii) coarse-scale (>100 m²) surface reflectance from hundreds of hectares of heavily infested liana forest stands in French Guiana and Bolivia. Using these data, we find consistent spectral signal of liana-infested canopies across sites with a mean inter-site correlation of 89% (range 74-94%). However, as we find no consistent difference between liana and tree leaves, a distinct liana spectral signal appears to only manifests at the canopy and stand scales (>1m²). To better understand this signal, we implement mechanistic radiative transfer models capable of modeling the vertically stratificatied non-linear mixing of spectral signals intrinsic to lianas infestation of forest canopies. Next, we inversely fit the models to observed spectral signals of lianas at all scales to identify key biochemical or biophysical processes. We then corroborate our model results with field data on liana leaf chemistry and canopy structural properties. Our results suggest that a liana-specific spectral distribution arises due to the combination of cheaply constructed leaves and efficient light interception. A model experiment revealed that the spectral distribution was most sensitive to lower leaf and water mass per unit area, affecting the absorption of NIR and SWIR radiation, and a more planophile (flatter) leaf angle distribution. Finally, we evaluate the theoretical discernibility of lianas from trees and how this varies with remote sensing platforms and resolution. We end by discussing the potential, limitations and risks of applying automated classifiers to detect lianas from remotely sensed data at large scales.