Accuracy of remote sensing and machine learning estimation of grassland aboveground biomass: a systematic review and meta-regression - study protocol and pre-registration

This is the pre-registered protocol (version 1) for a systematic review and meta-regression of the accuracy of grassland aboveground biomass (AGB) estimation from remotely sensed data. Three previous syntheses (Maake et al., 2023, Geomatics 3(4):478-500; Morais et al., 2021, Ecological Indicators 130:108081; Velazquez et al., 2026, Remote Sensing 18(19):3401) described this literature but performed no pooled effect size estimation or meta-regression.Velázquez et al. (2026) noted qualitatively that random hold-out or k-fold validation can inflate R-squared when field samples are spatially autocorrelated — which is almost always the case for grassland sampling — but did not quantify this effect. This review asks how much of the reported accuracy of grassland AGB remote sensing estimation reflects study design choices rather than genuine predictive skill. Pre-registered hypotheses: H1 (primary), studies using non-independent validation report significantly lower nRMSE(%) than those using independent or spatially blocked validation, indicating apparent accuracy inflation; H2, the apparent advantage of machine learning over traditional regression shrinks after controlling for validation design; H3, severe plot-to-pixel scale mismatch increases reported error; H4, smaller studies report apparently better accuracy, consistent with a small-sample bias. Primary effect size: log(nRMSE%). Secondary: Fisher z-transform of R-squared. Synthesis by multi-level random-effects meta-regression (metafor::rma.mv in R, random = ~1|study_id/row_id). Condition for quantitative synthesis: at least 60 extractable effect sizes and at least 10 studies with independent or spatially blocked validation; otherwise downgraded to a descriptive systematic review. This record was deposited before the commencement of full-scale data extraction.