Mapping tree species diversity across the Amazon using remote sensing, diverse environmental data and machine learning

International Journal of Applied Earth Observation and Geoinformation, vol. 148, pp. 105226

Abstract

• Integrated remote sensing and environmental data for Amazon tree diversity mapping. • Extreme gradient boosting identified as the best-performing algorithm. • Climate, soil and vegetation are influential, complementary predictors. • Generated 1 km resolution maps, showing spatial patterns at high resolution. Tropical forests are biodiversity hotspots facing increasing threats from climate change and anthropogenic pressures. Tree species diversity is a key indicator of forest biodiversity, making accurate and spatially detailed information essential for effective monitoring and conservation. For large-scale tree species diversity mapping, satellite data and machine learning offer great potential, but challenges remain, including limited field data, coarse spatial resolution, and complex relationships between in-situ diversity and remotely sensed metrics. This study investigates the potential of integrating remote sensing, environmental data and machine learning for tree species diversity mapping across the Amazon. Specifically, we 1) compared performances of widely used algorithms (random forest, extreme gradient boosting, artificial neural networks, support vector regression and ordinary least squares), 2) identified key environmental predictors, and 3) produced 1-km resolution maps to assess their spatial patterns. Our results show that extreme gradient boosting outperformed other algorithms. Climate and soil emerged as the most influential drivers of broad-scale diversity patterns, while remote sensing metrics were also influential in adding fine-scale spatial patterns. High tree diversity was associated most with climate stability, high precipitation, soil characteristics, and spatial heterogeneity of vegetation indices. The resulting maps aligned well with field observations and prior studies, while showing the potential of fine-scale mapping with competitive accuracy and low uncertainty using remotely sensed data. These findings thus demonstrate the effectiveness of integrating remote sensing, environmental data and machine learning for tree species diversity mapping in tropical forests, in support of biodiversity assessments and conservation measures.

Authors 5

1. Shoyo Nakamura corresponding

   Wageningen University & Research

   Affiliation as printed

   Earth Systems and Global Change Group, Wageningen University & Research, Droevendaalsesteeg 3, 6708 PB Wageningen, the Netherlands

2. Nandin‐Erdene Tsendbazar

   Wageningen University & Research

   Affiliation as printed

   Laboratory of Geo-Information Science and Remote Sensing, Wageningen University & Research, Droevendaalsesteeg 3, 6708 PB Wageningen, the Netherlands

3. Hans ter Steege Aachen Naturalis Biodiversity Center

   Naturalis Biodiversity Center · Utrecht University

   Affiliation as printed

   Naturalis Biodiversity Center, PO Box 9517, 2300 RA Leiden, the Netherlands

   Quantitative Biodiversity Dynamics, Utrecht University, Padualaan 8, 3584 CH Utrecht, the Netherlands

4. Lars Hein

   Wageningen University & Research

   Affiliation as printed

   Earth Systems and Global Change Group, Wageningen University & Research, Droevendaalsesteeg 3, 6708 PB Wageningen, the Netherlands

5. Jannik Schultner

   Wageningen University & Research

   Affiliation as printed

   Earth Systems and Global Change Group, Wageningen University & Research, Droevendaalsesteeg 3, 6708 PB Wageningen, the Netherlands