Cerrado Pilot: Short-term Land-Use Transition Modeling from MapBiomas Time Series under Latent Conversion and Regeneration Processes (1985–2024)

This pilot aims to model short-term land-use dynamics in the Brazilian Cerrado (targets: 2027 and 2030) using MapBiomas annual land-use/land-cover (LULC) maps (1985–2024) and auxiliary biophysical and legal-context layers. A key challenge is that MapBiomas annual class changes can compress gradual processes into apparent one-year transitions. In particular, (i) conversion from native vegetation to intensive agriculture may appear as sequences such as N→P→S (native vegetation to planted pasture to soy), where intermediate years can reflect conversion stages rather than stable land use; and (ii) regeneration from use to native vegetation (P/M→N) may appear as abrupt one-year transitions despite being ecologically gradual and potentially preceded by detectable declines in pasture condition. We therefore treat conversion and regeneration as latent, multi-year processes and explicitly quantify the duration and uncertainty of intermediate states. We will estimate an empirical temporal threshold k from the distribution of pasture persistence prior to agriculture (e.g., N→(P for d years)→S) and handle MapBiomas “Mosaic of Uses” (class 21), which can arise as a residual of integration rules, as an “unknown/low-confidence” state through masking and alternative imputation strategies. For regeneration, we will test whether pre-transition pasture condition (e.g., vigor, GPP, structural proxies) and their recent trends improve prediction of P/M→N events. The pilot will produce a reproducible dataset and evaluation protocol for spatially and temporally robust short-term forecasting in the Cerrado, and a set of design choices to inform subsequent model architecture development.