LARGE-SCALE VEGETATION HEIGHT MAPPING FROM SENTINEL DATA USING DEEP LEARNING Anders U. Waldeland, Arnt-Børre Salberg, Øivind D. Trier Norwegian Computing Center Department SAMBA P.O. Box 114 Blindern, N-0314 OSLO Andreas Vollrath ESA - European Space Agency ESRIN Largo G. Galilei 1, 00044 Frascati (RM), Italy ABSTRACT The deep learning revolution in computer vision has enabled a potential for creating new value chains for Earth observa- tion that significantly enhances the analysis of satellite data for tasks like land cover mapping, change analysis, and ob- ject detection. We demonstrate a deep learning based value chain for the task of mapping vegetation height in the Liwale region in Tanzania using Sentinel-1 and -2 data. As ground truth data we use lidar measurements, which are processed to provide the average vegetation height per Sentinel-2 pixel grid (10 m). We apply the UNet, which is a widely used neu- ral network for segmentation tasks in computer vision, to es- timate average vegetation height from the Sentinel data. Pre- liminary results show that we are able to map the forest extent with high accuracy, with an RMSE of 3.5 m for Sentinel-2 data and 4.6 m for the Sentinel-1 data. Index Terms— vegetation height, convolutional neural network, Sentinel-1/2 1. INTRODUCTION Vegetation height may be used to characterize the structure of a forest. It is known to correlate with important biophysical parameters like primary productivity, above-ground biomass, and bio-diversity [1, 2]. 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 provide canopy height over ground maps densely and accurately, but the cost and the limited area covered makes it infeasible for large-scale monitoring. Trier et al. [1] estimated vegetation height from Land- sat data covering the Liwale area in Tanzania. A regression model between the average vegetation height computed from the lidar data and the specific leaf area vegetation index com- puted from the Landsat data, was established. By using all available Landsat acquisitions of the same area within 1 year, Thanks to ESA - European Space Agency for funding. and producing a yearly estimate of vegetation height, the esti- mation error variance was reduced. The variance was further reduced by Kalman filtering the sequence of yearly estimates. Lang et al. [2] also estimated the vegetation height, but from Sentinel-2 data. Their study areas were Gabon and Switzerland, and their apporach was to train a deep convolu- tional neural network (CNN) to regress per-pixel vegetation height. Their results showed good qualitative agreement with existing vegetation height maps, and the authors demon- strated that vegetation height maps with 10 m pixel-spacing can be derived at country scale from Sentinel-2 imagery. As stated by Lang et al. [2], single-pixel based prediction of vegetation height at 10m pixel spacing is not suitable due to physical phenomena like shadowing, roughness, and species distribution that extends across neighboring pixels. Deep CNN architectures like UNet are perfectly tailored to account for the spatial context of the problem. In this study, we explore and compare Sentinel-1 and - 2 data to estimate the height of dry tropical vegetation. We consider the same area in as Trier et al. [1], but establishes a regression model using a deep CNN to estimate the vegeta- tion height from the Sentinel data. The network is based on the UNet [3] architecture, working in regression mode, and implemented in the deep learning framework for large-scale processing of Sentinel data proposed by Salberg and Walde- land [4]. 2. STUDY AREA AND DATA 2.1. Study area The study area was Liwale in Tanzania (S9 ◦ 54’, E37 ◦ 38’). It covers 15,867 km 2 of Miombo woodlands with altitudes in the range 150–900 m above sea level. The rainfall pattern in Liwale is bi-modal with a dry season from June to Octo- ber. A short rainy period usually starts in late November and lasts until January. Normally, there is a dry spell in February followed by a longer wet season that lasts from March until May.