Country-wide high-resolution vegetation height mapping with Sentinel-2

Highlights

• Vegetation height at 10 m ground sampling distance is regressed from Sentinel-2

• Country-wide maps are computed for Switzerland and Gabon

• Mean absolute error (MAE) of 1.7 m in Switzerland and 4.3 m in Gabon

• The deep convolutional neural network correctly predicts vegetation heights up to 50 m

Abstract

Sentinel-2 multi-spectral images collected over periods of several months were used to estimate vegetation height for Gabon and Switzerland. A deep convolutional neural network (CNN) was trained to extract suitable spectral and textural features from reflectance images and to regress per-pixel vegetation height. In Gabon, reference heights for training and validation were derived from airborne LiDAR measurements. In Switzerland, reference heights were taken from an existing canopy height model derived via photogrammetric surface reconstruction. The resulting maps have a mean absolute error (MAE) of 1.7 m in Switzerland and 4.3 m in Gabon (a root mean square error (RMSE) of 3.4 m and 5.6 m, respectively), and correctly estimate vegetation heights up to >50 m. They also show good qualitative agreement with existing vegetation height maps. Our work demonstrates that, given a moderate amount of reference data (i.e., 2000 km² in Gabon and ≈5800 km² in Switzerland), high-resolution vegetation height maps with 10 m ground sampling distance (GSD) can be derived at country scale from Sentinel-2 imagery.

Introduction

Vegetation height is a basic variable to characterise a forest's structure, and is known to correlate with important biophysical parameters like primary productivity (Thomas et al., 2008), above-ground biomass (Anderson et al., 2006) and bio-diversity (Goetz et al., 2007). However, direct measurement of tree height does not scale to large areas and/or high spatial resolution: in-situ observations are in practice only feasible for a limited number of sample plots and logging sites. Airborne light detection and ranging (LiDAR) can map canopy height over ground densely and accurately, but the financial cost and the limited area covered per day only allow for small regional projects (some countries of moderate size have complete coverage, but with low revisit times of several years between subsequent acquisitions). Finally, space-borne LiDAR provides world-wide coverage, but the measurements are sparse in both space and time: distances between adjacent profiles are in the tens of kilometers, and nearby observations have been acquired up to 6 years apart. After 7 years of data collection, the point density in Gabon, for example, is only 1.26 shots per km² (Baghdadi et al., 2013). Moreover, each measurement is averaged over a ground footprint of 70 m radius.

Hence, dense wide-area maps of canopy height are typically obtained by regression from multi-spectral satellite images, using in-situ or LiDAR heights as reference data to fit the regression model (Lefsky, 2010, Hudak et al., 2002). This approach has made it possible to produce tree height maps with ground resolutions down to 30 m, by exploiting the Landsat archive (Hansen et al., 2016).

Here, we demonstrate country-wide mapping of canopy height with a ground resolution of 10 m, by regression from Sentinel-2 multi-spectral data. At such high resolutions, the spectral signature of an individual pixel is no longer sufficient to predict tree height. Rather, the physical phenomena underlying the monocular prediction of tree height, like shadowing, roughness, and species distribution give rise to reflectance patterns across neighbourhoods of multiple pixels. It is, however, not obvious how to encode the resulting image textures into predictive feature descriptors that support the regression. To sidestep this problem, we resort to deep learning. Recent progress in computer vision and image analysis has impressively demonstrated that very deep¹ convolutional neural networks (CNNs) are able to learn a tailored multi-level feature encoding for a given prediction task from raw images, given a sufficient (large) amount of training data. Our experiments reveal that texture patterns are particularly important in areas of high (tropical) forest, extending the sensitivity of the regressor to heights up to ≈55 m. End-to-end learning of rich contextual feature hierarchies underlies several successes of image and raster data analysis, including visual recognition of objects (Krizhevsky et al., 2012), understanding human speech from spectrograms (Abdel-Hamid et al., 2014) and assessment of positions in board games like go or chess (Silver et al., 2018).

We employ a deep convolutional neural network to regress country-wide canopy height for Gabon and Switzerland from 13-channel Sentinel-2 Level 2A images (corrected to bottom-of-atmosphere reflectance), using reference values obtained from airborne LiDAR scans and photogrammetric stereo matching as training data. The two countries were selected because in both we have access to reference data for training and quantitative evaluation: in Switzerland from the national forest inventory program; in Gabon via NASA's LVIS project. At the same time, the two countries are very different in terms of their geography and biomes, which supports our belief that the proposed approach can be scaled up to global coverage. Importantly, we also find that no long time series or multi-temporal signatures are required. A few observations per pixel (4 to 12) already achieve low prediction errors – in fact, even predicting from a single image yields fairly decent results. This means that, at the 5-day revisit cycle of Sentinel-2, we are able to obtain almost complete coverage using only the 10 clearest images within the leaf-on season (May–September) for Switzerland or within a period of 12 months in tropical forest regions with frequent cloud cover.

Our work is, to our knowledge, the first to demonstrate large-scale vegetation height mapping from optical satellites at 10 m GSD. The model is able to retrieve tree heights up to ≈55 m, well beyond the saturation level of existing high-resolution canopy height maps (e.g., Hansen et al., 2016). At the technical level, we are not aware of any other work that employs deep CNNs for canopy height estimation from optical satellite data.

Based on the present work, the next goal is to generate a global, wall-to-wall map of canopy height.