Castor Beans Identification in Sugarcane Plantations Using Active Learning

• Juan Carlos Conceição de Lima Sales UFMS http://orcid.org/0000-0002-7590-6275
• Mário de Araújo Carvalho UFMS
• José Marcato Junior UFMS
• Wesley Nunes Goncalves UFMS

Resumo

Castor beans have multiple applications but can become weeds in several crops. Manual identification is unfeasible in large plantations. Machine learning models can make identification efficient but require a lot of labeled data. Many images of a plantation would be similar and wouldn't improve the model's performance. Active learning (AL) allows labeling only relevant data, often surpassing the performance of models trained on entire datasets. In this work, we detect castor beans, with an AL method that uses self-supervised pretext tasks to separate data for labeling. Models based on pretext tasks presented a decrease in recall relative to the model trained in the whole dataset, which has 93%. We also trained a pseudo-task that separates data with a reasonable concentration of castor bean images. The pseudo-task classifier obtained a 92% recall, being trained in less than 1% of the dataset.

Palavras-chave: Active Learning, Machine Learning, Weed Control, Castor Beans