From Remote Sensing Accuracy to MRV Readiness: A Systematic Review and GeoAI Framework for Tropical Forest Biomass Estimation

Remote sensing combined with machine learning (ML) and deep learning (DL) is widely used to estimate tropical forest aboveground biomass (AGB), yet it is unclear how fully published studies document the evidence that Measurement, Reporting and Verification (MRV) of forest carbon would require. We assessed this reporting evidence using a five-dimension framework (accuracy reporting, uncertainty quantification, spatial validation, temporal consistency and transparency) applied to a systematically identified set of empirical studies. From 1706 records retrieved from Scopus and Lens.org on 4 March 2026, 1146 unique works were screened against a locked eligibility codebook; 153 reports were assessed in full text where accessible, and 59 empirical ML/DL remote-sensing AGB studies in tropical or subtropical forests formed the core review set. For each dimension we coded evidence as confirmed, reported as absent, insufficient or inaccessible, and report the share of studies between the confirmed count and the confirmed-plus-undetermined count. Accuracy metrics were reported in 66–88% of studies, but spatially independent validation was confirmed in only 32–42%, spatially explicit or area-level uncertainty in 44–54%, multi-date analysis with an explicit consistency method in 14–22%, public code in 14–29% and public reference data in 15–29%. Only 12 studies confirmed both independent validation and propagated uncertainty. The results describe the evidence reported in research papers and do not assess compliance of any product with a carbon standard. Screening and evidence extraction were AI-assisted against a fixed codebook with a verbatim evidence phrase for every decision, and a stratified sample of 60 screening decisions was independently verified by two co-authors (agreement 88% and 78%). We recommend that biomass studies intended to inform MRV report spatially blocked or independent-site validation, propagate uncertainty to area totals, and release code and reference data.