As part of the EU Common Agricultural Policy (CAP) reform of 2020, each EU member country is expected to suggest new farmland management protocols. Currently, farmers must manually declare each year their crop types into the Land-Parcel Identification Systems (LPIS), a geographic information system identifying the land use of agricultural parcels within each EU member country. Such a protocol remains tedious and error-prone. Automatic Earth observation image analysis can help achieving such a task. Leveraging the recent availability of precise and frequent Sentinel acquisitions, this work aims to automate the LPIS update. We propose modeling the crop type of parcels from a sequence of (radar and optical) satellite acquisitions, as well as LPIS entries of previous years, with a linear-chain Conditional Random Field. The novelty lies on the fusion of multi-modal images at the feature level and the integration of temporal knowledge extracted from existing land-cover databases. We tested our model on two large-scale French study areas (≥1250 km 2), which are geographically distant and show different agronomic rules: the Seine et Marne (North of France) and the Alpes de Haute-Provence (South East). We use a granular nomenclature comprised of 25 categories. Our model demonstrates promising results for the task of automating the LPIS update: 89.0% overall accuracy is reached in Seine et Marne (10 categories of the 25 present on the area) and 72.9% in Alpes de Haute-Provence 27 (14 categories). We show that the temporal modeling increases the accuracy by +2.6% and +4.6%, respectively.