Active Learning for Image Classification: A Comprehensive Analysis in Agriculture

Precision agriculture allows for the sustainable improvement of agricultural products by introducing technologies that provide crop-specific data, which supervised algorithms can process. However, supervised algorithms require expert-labeled data, which can be highly expensive in agricultural applications. Given this problem, active learning (AL) arises as an alternative to reduce the need to annotate training examples manually. This paper analyzes the use of AL in classifying agricultural crop images. To evaluate AL, two datasets with information on fruit and vegetable images were used on two neural network-based algorithms. The classification results indicate that AL reduced the number of training examples to achieve a given performance. Additionally, the pseudo-labels of the supervised algorithms, a stopping criterion, and the explainability of the predictions were analyzed. These analyses allowed to assess the applicability of AL in agriculture to understand the learning process of the supervised algorithms.