Scholarlink

<p>In this study, we present the integration of tracker expertise with artificial intelligence (AI) for wildlife species classification and its application to human-in-the-loop with an investigation in explainable AI. We collected images of wildlife tracks, built AI models from the track images, classified species based on the best performing model, and expert trackers evaluated the results against the AI model. The wildlife species included black rhinoceros (<i>Diceros bicornis</i>), blue wildebeest (<i>Connochaetes taurinus</i>), giraffe (<i>Giraffa camelopardalis</i>), and white rhinoceros (<i>Ceratotherium simum</i>). Two expert trackers and one non-expert tracker ranked the image quality of 3039 tracks. We then trained the AI model with different number of training images per class using different hyperparameter settings. The best-performing AI model was chosen and evaluated. Afterwards, 36 expert trackers evaluated the resulting model: one set using raw images only, and another set using raw with heatmap images. For the same hyperparameter settings with the best model performance evaluation, our method considerably increased the mean average precision@50–95 by 10.42% against a non-expert tracker. In addition, the required number of images for model training can be reduced by 25% when using inputs from a highly skilled expert tracker. The visual heatmaps provided a means in performing explainable AI by visually presenting the features of the tracks based on colours that help to guide the expert tracker evaluation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improving wildlife track classification through human-in-the-loop method and explainable AI

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月20日

查看全文
  • Tinao Petso,
  • Rodrigo S. Jamisola Jr.,
  • Sky Alibhai,
  • Montiredi Lebopo,
  • Ketshabile Peter,
  • Molaletsa Namoshe,
  • Wazha Mmereki,
  • Zoe Jewell

摘要

In this study, we present the integration of tracker expertise with artificial intelligence (AI) for wildlife species classification and its application to human-in-the-loop with an investigation in explainable AI. We collected images of wildlife tracks, built AI models from the track images, classified species based on the best performing model, and expert trackers evaluated the results against the AI model. The wildlife species included black rhinoceros (Diceros bicornis), blue wildebeest (Connochaetes taurinus), giraffe (Giraffa camelopardalis), and white rhinoceros (Ceratotherium simum). Two expert trackers and one non-expert tracker ranked the image quality of 3039 tracks. We then trained the AI model with different number of training images per class using different hyperparameter settings. The best-performing AI model was chosen and evaluated. Afterwards, 36 expert trackers evaluated the resulting model: one set using raw images only, and another set using raw with heatmap images. For the same hyperparameter settings with the best model performance evaluation, our method considerably increased the mean average precision@50–95 by 10.42% against a non-expert tracker. In addition, the required number of images for model training can be reduced by 25% when using inputs from a highly skilled expert tracker. The visual heatmaps provided a means in performing explainable AI by visually presenting the features of the tracks based on colours that help to guide the expert tracker evaluation.