Vegetation variables such as leaf area index (LAI) and leaf chlorophyll content (Cab) are important inputs for vegetation growth models. LAI and Cab can be estimated from remote sensing data using either empirical or physically-based approaches. The latter are more generally applicable because they can easily be adapted to different sensors, acquisition geometries, and vegetation types. They estimate vegetation variables through inversion of radiative transfer models. Such inversions are ill-posed but can be regularized by coupling models, by using a priori information, and spatial and/or temporal constraints. Striving to improve the accuracy of LAI and Cab estimates from single remote sensing images, this contribution proposes a Bayesian object-based approach to invert at-sensor radiance data, combining the strengths of regularization by model coupling, as well as using a priori data and object-level spatial constraints.The approach was applied to a study area consisting of homogeneous agricultural fields, which were used as objects for applying the spatial constraints. LAI and Cab were estimated from at-sensor radiance data of the Airborne Prism EXperiment (APEX) imaging spectrometer by inverting the coupled SLC-MODTRAN4 canopy–atmosphere model. The estimation was implemented in two steps. In the first step, up to six variables were estimated for each object using a Bayesian optimization algorithm. In the second step, a look-up-table (LUT) was built for each object with only LAI and Cab as free variables, constraining the values of all other variables to the values obtained in the first step. The results indicated that the Bayesian object-based approach estimated LAI more accurately (R2 = 0.45 and RMSE = 1.0) than a LUT with a Bayesian cost function (LUT–BCF) approach (R2 = 0.22 and RMSE = 2.1), and Cab with a smaller absolute bias (− 9 versus − 23 μg/cm2).The results of this study are an important contribution to further improve the regularization of ill-posed RT model inversions. The proposed approach allows reducing uncertainties of estimated vegetation variables, which is essential to support various environmental applications. The definition of objects and a priori data in cases where less extensive ground data are available, as well as the definition of the observation covariance matrix, are critical issues which require further research.