Satellite-based Machine Learning modelling of Ecosystem Services indicators: A review and meta-analysis

Satellite-based Machine Learning (ML) modelling has emerged as a powerful tool to understand and quantify spatial relationships between landscape dynamics, biophysical variables and natural stocks. Ecosystem Services indicators (ESi) provide qualitative and quantitative information aiding the assessment of ecosystems’ status. Through a systematic meta-analysis following the PRISMA guidelines, studies from one decade (2012–2022) were analyzed and synthesized. The results indicated that Random Forest emerged as the most frequently utilized ML algorithm, while Landsat missions stood out as the primary source of Satellite Earth Observation (SEO) data. Nonetheless, authors favoured Sentinel-2 due to its superior spatial, spectral, and temporal resolution. While 30% of the examined studies focused on modelling proxies of climate regulation services, assessments of natural stocks such as biomass, water, food production, and raw materials were also frequently applied. Meta-analysis illustrated the utilization of classification and regression tasks in estimating measurements of ecosystems' extent and conditions and findings underscored the connections between established methods and their replication. This study offers current perspectives on existing satellite-based approaches, contributing to the ongoing efforts to employ ML and artificial intelligence for unveiling the potential of SEO data and technologies in modelling ESi. •ESi provide qualitative and quantitative information for ecosystems' status assessment.•The prediction of carbon storage was the goal of 30% of the ML models developed.•Climate regulation, Habitat, and genetic resources were the ESi most modelled.•Even in regression models, classification tasks were commonly the first step taken.•Binary classifications were employed to map ecosystems' extent.