Scale matters: Spatial resolution impacts tropical leaf phenology characterized by multi-source satellite remote sensing with an ecological-constrained deep learning model

Guangqin Song, Jing Wang, Yingyi Zhao, Dedi Yang, Calvin K.F. Lee, Zhengfei Guo, Matteo Detto, Bruna Alberton, Patricia Morellato, Bruce Nelson, Jin Wu

Research output: Contribution to journalArticlepeer-review

8 Scopus citations

Abstract

Accurate monitoring of tropical leaf phenology, such as the leaf-on/off status, at both individual and ecosystem scales is essential for understanding and modelling tropical forest carbon and water cycles, and their sensitivity to climate change. The discrepancy between tree-crown size and pixel size (i.e., spatial resolution) across orbital sensors can affect the capability of cross-scale phenology monitoring, an aspect that remains understudied. To examine the impact of spatial resolution on tropical leaf phenology monitoring, we applied a spectral index-guided, ecologically constrained autoencoder (IG-ECAE) to automatically generate a deciduousness metric (i.e., percentage of upper canopy area that is leaf-off status within an image pixel) from simulated VIS-NIR PlanetScope data at a range of resolutions from 3 m to 30 m, as well as from VIS-NIR data of three satellite platforms with the same range of spatial resolutions (3 m PlanetScope, 10 m Sentinel-2, and 30 m Landsat-8). We compared the deciduousness metrics derived from the simulated and satellite data to corresponding measurements derived from WorldView-2 (three sites) and local phenocams (four sites) at five tropical forest sites. Our results revealed that: (1) the IG-ECAE model captured the amount of deciduousness across spatial scales, with the highest accuracy obtained from PlanetScope, followed by Sentinel-2 and Landsat-8; (2) coarser spatial resolutions led to lower accuracies in tropical deciduousness monitoring, as demonstrated by both simulated PlanetScope data across various spatial resolutions and real satellite data; and (3) while not as accurate in capturing fine-scale tropical phenological diversity as PlanetScope, Sentinel-2 provided satisfactory monitoring of deciduousness seasonality at the ecosystem level consistently across all phenocam sites, whereas Landsat-8 failed to do so. Collectively, this study provides a robust assessment for advancing cross-scale tropical leaf phenology monitoring with potential for extension to pan-tropical regions and highlights the impact of spatial resolution on such monitoring efforts.

Original languageEnglish
Article number114027
JournalRemote Sensing of Environment
Volume304
DOIs
StatePublished - Apr 1 2024
Externally publishedYes

Keywords

  • Deep learning
  • Ecosystem deciduousness
  • Leaf phenology
  • Phenological diversity
  • Satellite remote sensing
  • Spatial resolution
  • Spectral unmixing
  • Tropical forest

Fingerprint

Dive into the research topics of 'Scale matters: Spatial resolution impacts tropical leaf phenology characterized by multi-source satellite remote sensing with an ecological-constrained deep learning model'. Together they form a unique fingerprint.

Cite this