AutoML for estimating grass height from ETM+/OLI data from field measurements at a nature reserve. Issue 1 (31st December 2022)

• Record Type: Journal Article
• Authors: César de Sá, Nuno; Baratchi, Mitra; Buitenhuis, Vincent; Cornelissen, Perry; van Bodegom, Peter M.
• Is Part Of: GIScience & remote sensing. Volume 59:Issue 1(2022)
• Page Start: 2164
• Page End: 2183
• Publication Date: 2022-12-31
• Subjects: AutoML -- machine learning -- remote sensing -- grassland -- Landsat; Geodesy -- Periodicals; Cartography -- Periodicals; Aerial photogrammetry -- Periodicals; Remote sensing -- Periodicals; 526.05
• DOI: 10.1080/15481603.2022.2152304
• Languages: English
• ISSNs: 1548-1603

Abstract

ABSTRACT: Remote sensing (RS) is now a standard tool used for grassland monitoring thanks to the availability of data at an unprecedented spatial and temporal resolution. The approaches to monitor grasslands often rely on the use of vegetation indices (e.g. NDVI) and empirical models trained on field data collected in tandem with the RS data. The best combination of models and features is often found by ad-hoc experimentation by the expert. This "classic" approach does not necessarily result in the best possible model. Automatic machine learning (AutoML) allows to automate this procedure by identifying the best possible pipeline in a data-driven manner. This study assesses the applicability of two distinct AutoML algorithms – AutoSklearn and AutoGluon – to monitor grass height from RS data and to systematically compare them to "classic" RS approaches. Grass height was estimated from Landsat ETM+ and OLI for a well-known conservation area as a case study. The "classic" RS approach emulated all possible ad hoc decisions by comparing all combinations of bands and vegetation indices against a naive use of the AutoML systems. While model selection and optimization are automated within AutoML models, for the "classic" RS approach, we used Bayesian optimization for hyperparameter tuning. We found that AutoML methods outperformed "classic" methods with the test error varying between ~1.73 cm ± 0.02 and ~1.78 cm ± 0.03 while for the "classic" methods it varied between ~1.84 cm ± 0.03 and ~2.81 cm ± 0.02. In the case of the "classic" methods, our exhaustive exploration of the possible feature combinations showed that while vegetation indices were always selected for the best models, which index got selected depended on the algorithm. The performance of AutoML compared to "classic" RS approaches clearly demonstrates the ability of these methods to quickly and effectively identify high-performing models. However, as this work focused on a single case-study, the results cannot be directly generalized to other study areas. Nevertheless, it provided a number of insights into future research opportunities to improve the use of AutoML in RS.