Role of Auxiliary Variables as Inputs or Targets in Spatial Generalization of Satellite Image Time Series Foundation Models

Recent studies report that Earth Observation (EO) Foundation Models (FMs) using satellite image time series (SITS) lose performance when tested on unseen regions. Integrating auxiliary variables during training has shown promise in mitigating this performance drop. A first line of work in the development of FMs integrates these auxiliary variables as input while other works propose to use them as targets. To our knowledge, the two integrations have never been compared under the same conditions such as architecture, pre-training strategy and dataset. This study examines whether integrating auxiliary variables as inputs or as pre-training targets changes in-domain and cross-domain performance. We perform the evaluation on two tasks: tree and crop species mapping, using SEMIS, an encoder we designed to extract spatial, spectral and temporal features. We find that the selected auxiliaries used as inputs improve in-domain performance (+7.0/+0.6 macro-F1 for trees and crops) and reduce cross-domain performance (-1.6/-1.4), while our results suggest using auxiliaries as targets improves in-domain on trees (+4.6) but not on crops (-0.9) and improves cross-domain performance in both tasks (+2.6/+0.4). Targets are therefore the preferable integration when the objective is transfer: they outperform input integration in cross-domain in all considered regions, and they require no auxiliary data at inference.