A Map Should Show Uncertainty

A classification map usually looks more certain than the evidence behind it. Every pixel receives a colour, every parcel receives a label, and the visual result can hide the fact that the model was trained on a small number of fields, that the current season is unusual, or that a large part of the time series was reconstructed rather than observed. In orchard work, uncertainty is not a failure to be removed from the interface; it is part of the information a grower needs before spending time in the field.

There are several ways to make that uncertainty useful. A model can return calibrated class probabilities instead of only the most likely label. A parcel result can include valid-observation coverage, the proportion of changed pixels, and how far the current curve sits from the reference distribution. A national-scale orchard mapping study in Italy combined a probabilistic XGBoost classifier with Bayesian calibration using prior information from national statistics, which is a practical example of treating probabilities as something that should be checked rather than assumed. Hierarchical classification provides another useful option: when the fine class is uncertain, the system can still return a more reliable broad class such as “orchard.” See Lodato et al. and the hierarchical time-series study.

Uncertainty also changes the design of validation. Randomly splitting neighbouring pixels can make a model appear stronger than it is because the training and test sets contain nearly identical pieces of the same field. A more honest test holds out whole fields, seasons, or regions, then reports where performance falls. The result is not as flattering, but it tells the team whether the map can support a new block, a new year, or only the place where it was trained. Confidence is valuable when it narrows the next question; it is dangerous when it merely decorates a weak answer.