Generalization of image classification models using fair optimization techniques

Deep learning algorithms can achieve great accuracies, but there are also concerns about biases and generalizability to unseen data. Biases can arise for many reasons, one of them being that certain classes are better represented in training datasets than others. As the classifier sees examples of the majority class more than minority classes, it will also learn to recognize the majority class better than the minority classes. One way to reduce this bias is to use optimization functions that not only maximize the accuracy but also minimize the difference in performance amongst the different classes. The TensorFlow Constrained Optimization Library is one example of a tool to do this.

One possible use case is the identification of the minority crops cultivated in a diverse agricultural landscape. This use case is, however, not cast in stone. Minority crop class can be replaced by any other land cover/land use class of interest.