The crop- and land-group markers determine whether an agricultural parcel (FOI — feature of interest) is consistent with its declared crop and land use. Both markers share the same architecture; this article describes it using the crop-group marker as the example.

Why crops are grouped

Each FOI has a declared crop, and that crop defines how the parcel's signal time series behaves through the year. Some crops are easy to tell apart — corn and grass NDVI time series look vastly different. Others are difficult to recognize as separate with remote sensing: for example, in two sample parcels declared as winter wheat and winter barley, the two crops are pretty much inseparable (ignoring the secondary crop sown after July). Other such groups include grass and grass-like annual crops such as alfalfa, summer grains, and many more.

To achieve more accurate and reliable results, crops that behave alike are often grouped, and each group is treated as a single training/prediction class, although this is not strictly necessary. Treating each crop as its own class is possible in principle, but usually comes at a drastic cost in performance — especially confusion between crops within the same group. As a reference point: in Slovenia there are roughly 200 crops grouped into 31 crop groups, with grasses, winter grains, and summer grains among the most populous.

How the model is trained

The marker uses a Long Short-Term Memory (LSTM) model, a recurrent neural network architecture tailored to recognizing classes from patterns in time series such as satellite data.

Training data is based on the target year's agricultural parcel reference data (declarations). Historical data can also be included where reference data is available — this boosts underrepresented classes and improves accuracy on classes that are rare in a given year.

Cleaning the training data

Because declarations serve as the target classes, and declarations can be incorrect, likely-wrong claims are filtered out of the training set using other markers:

• Distance marker outliers — outliers from the distance marker (for example, a parcel declared as meadow where the distance and similarity markers point to the claim likely being incorrect).
• Homogeneity marker flags — parcels that are likely heterogeneous (several crops growing at once), which could confuse the model.

Five-fold prediction

After filtering, the FOIs are randomly divided into five folds (A–E). For each fold, a separate model is trained on the other four folds and then predicts on the left-out fold — for example, predictions for fold A come from a model trained on folds B–E. This way, every FOI in the dataset receives a prediction from a model that never saw it during training.

The land-group marker

The land-group marker uses exactly the same building blocks, but groups declared land-use labels into land-groups instead of crops. In addition to agricultural land uses, non-agricultural classes are included in training — most prominently forests and built-up land. Including them means the marker can recognize when an agricultural parcel's land use has transformed into a non-agricultural one.

To read more on this, please see the Crop and Land Use Markers article.