Canopy height estimation from PlanetScope time series with spatio-temporal deep learning

Abstract

Canopy height mapping is critical for assessing forest structure, forest resilience, carbon stocks, habitat, and biodiversity, all of which are threatened by changing climate and weather extremes. While current tools utilizing lidar (e.g., GEDI) and multispectral imagery (e.g., Landsat, Sentinel-2, airborne imagery) produce canopy height products, significant challenges remain, particularly in capturing fine-scale spatial details across large areas with high frequency. PlanetScope CubeSat imagery, with its 3 m spatial resolution and near-daily frequency, offers a unique opportunity to estimate woody plant structure by capturing fine-scale texture and temporal patterns that shift throughout the year. In this study, we adapted a 3D Spatio-Temporal Convolutional Neural Network (ST-CNN) to estimate canopy height at 3 m resolution, utilizing sequential PlanetScope time series over five months, summer Sentinel-1 radar imagery, and solar illumination layers as inputs. We generated a large and diverse reference database covering 2,296 sample scenes (each scene = 768 × 768 m, totaling ∼ 135,000 ha) using a semi-automatic labeling process that leverages 23 aerial lidar surveys conducted in California between 2016 and 2021. Trained on a random selection of 2,046 scenes, the accuracy assessment on the remaining 250 scenes demonstrates strong performance across various ecoregions, capturing 80.8% of the observed variance in live canopy height with a mean absolute error (MAE) of 3.6 m and a bias of -0.53 m compared with aerial lidar. Analysis of all 681 GEDI footprints over the same testing scenes estimates the MAE of 6.5 m, bias of -1.82 m, and R 2 of 0.58 for the GEDI L2A Vector Canopy Top Height RH98 product. The ST-CNN model accurately identifies heterogeneous canopy structures, and shows sensitivity to canopies reaching 50 to 60 m in height. We found a major contribution from the PlanetScope time series, compared to a single PlanetScope image, and marginal benefits of including Sentinel-1 and terrain-based solar irradiance layers to improve performance on dense canopies or diverse topography. Example applications demonstrate the ability to generalize to different years, maintaining consistent predictions between years and capturing changes in canopy height over a seven year period (2017–2023) within 400 plots representing regrowth, minimal change, selective logging, and clear cut areas. We also demonstrate improved canopy height estimation compared to existing products from Landsat (MAE = 8.41 m) and Sentinel-2 (MAE = 7.19 m). A visualization tool displays our data alongside existing products for the Sierra Nevada in 2022. The Planet ST-CNN model, using a 15-day PlanetScope satellite time series, offers a scalable approach for annual canopy height estimation in California, achieving a high level of detail, often down to the resolution of individual trees. This improved capability of forest structure monitoring is expected to deliver crucial, comprehensive data for assessing and tracking forest carbon, biodiversity, and vulnerability, ultimately facilitating data-driven strategies to improve forest resilience. • A 3 × 3 m canopy height estimation workflow is developed combining optical and radar satellite data. • A semi-automatic workflow generates labels from aerial lidar and NAIP imagery. • A 3D convolutional neural network extracts spatio-temporal features from PlanetScope time series. • Scalability and robustness are demonstrated across California’s diverse ecoregions. • The model and data support annual high resolution forest structure assessment.

Topics & Concepts