Optimizing crop type mapping for fairness

Ensuring fairness in machine learning applications is critical, yet it remains underexplored in crop type mapping. While the consequences of imbalanced classes for supervised classification tasks are known to the field of Earth Observation, assessing classification results for sub-groups of societally sensitive attributes, such as parcel size in crop mapping, have received little attention. To address this gap, we evaluate established class imbalance correction methods: Random Oversampling (RO), Weighted Cross Entropy (WCE), and Focal Loss (FL); and two novel approaches that target both class imbalance and the performance disparity between small and large parcels: Random Oversampling with Resampling (RO-R), and Double Objective Weighted Cross Entropy (DOWCE). RO-R increases the representation of smaller parcels by redistributing random samples, whereas DOWCE applies higher penalties to the misclassification of smaller parcels. Hybrid methods (RO-DOWCE and RO-FL) were also evaluated. To assess their generalizability under varying conditions, the methods were tested on ten diverse datasets subsampled from the BreizhCrops dataset, covering Brittany, France. Results showed that RO-DOWCE was the most effective method at addressing class imbalance across the datasets, though not significantly different from RO-R, RO, and RO-FL. Additionally, cost-sensitive methods were generally less efficient at addressing class imbalance than sample balancing and hybrid approaches. These findings illustrate how broader discussions on Responsible AI and fairness are relevant for Earth Observation applications such as crop type mapping. Furthermore, the strategies to increase fairness presented here can be applied to classification tasks outside the domain of crop mapping.