Development of image classification models for the identification of earthworms exposed to glyphosate-based herbicide: a pilot study

• Autor(a) principal: de Batista, Diovana Gelati
• Data de Publicação: 2024
• Outros Autores: Pinheiro, Juliana Furlanetto, Ourique, Isadora Sulzbacher, Fernandes, Vítor Basto, Frantz, Rafael Z., Costa, Nuno, Heck, Thiago Gomes
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: https://doi.org/10.48797/sl.2024.217
• topic: Poster

Background: Glyphosate-based herbicides (GBH) may threaten ecosystems and human health [1]. Animal models using earthworms as environmental bioindicators have been proposed [2], but they must be practical and cheaper [3]. Objective: We test if machine learning models of earthworm image classification can be used to identify GBH-exposed environments. Methods: 144 adults Eisenia andrei earthworms were divided into Control (water), GBH1.5, GBH3.0, and GBH6.0 groups (Roundup® Original DI, equivalent to 1.5, 3.0, and 6.0 L/ha). After 48 hours, each worm was photographed at least two times with a mobile camera (76-88 images/group). Random images were used to train models (85%) and separated for testing (15%). Also, we generated 20 artificial images (AI) variations of each original image (OI) using data augmentation techniques using imgaug library [4], reaching >1,600 images/group. Thus, we trained models six times each in Google’s Teachable Machine with 50, 20, and 10 epochs (learning rate=0.001; batch size=16) using OI with the four (OI-4G) or two groups (OI-2G, Control vs. GBH6.0), or using AI (AI-4G or AI-2G). The resulting models were tested using Python with new images, and the accuracy was compared using 2-way ANOVA, followed by Tukey's test. Results: The OI-2G model showed better accuracy when trained with 50 epochs (P=0.02), but the AI-2G model presented the best accuracy in all epochs tested (P < 0.002). In contrast, the OI-4G model presented the worst performance compared to the others (P<0.0001) (% Accuracy: OI-4G=52±5; OI-2G=77±5; AI-4G=79±3; AI-2G=93±3). When tested, AI models had lower accuracy when compared to OI models (%Accuracy: OI-4G=47; OI-2G=86; AI-4G=38; AI-2G=65). Conclusions: It is possible to detect the presence of GBH in the soil by evaluating earthworm images using machine learning models, even with small sample sizes (photos) and without images created artificially. Models need to be improved to detect the concentration of GBH.