Deep Learning for Biodiversity Monitoring: Comparison of Custom Imaging Platforms and Deep Learning to Classify and Phenotype Earthworms

In a world where soil degradation is occurring at an alarming rate, studying organisms that can help restore it is essential. Earthworms play a crucial role in soil restoration and as indicators of overall soil health. They enhance soil nutrition and fertility, affect the presence of bacteria and fungi, and transform the soil structure, leading to aeration of the soil, among other things. Research indicates that earthworms can significantly assist in restoring soil in severely damaged areas, such as those affected by opencast mining, degraded Savanna soils, landfill soils, and saline soils. This potential is particularly promising for restoring agricultural soils impacted by overproduction. Classifying earthworms into functional groups (i.e., endogeic, epigeic, and aneic) is not the only way to assess their impact on soil. Access to taxonomic information about the earthworm species present allows for a more accurate estimation of soil quality since the ecological diversity of earthworms is key to maintaining soil health.
Deep learning models are highly effective at classification tasks and are making significant advancements in various fields, including medical image assessment, autonomous driving and insect classification. However, to my knowledge, there has only been carried very few studies that have applied deep learning for the automatic classification of earthworms. Therefore, the goal of this project is to classify earthworms into different genera and species, addressing a multiclass classification problem using deep learning models. This will be accomplished through three distinct work packages:
- Work Package 1 will focus on the classification of earthworms captured with mobile phone cameras using deep learning. This poses a challenge due to factors such as the low resolution and the way earthworms may fold when photographed.
- Work Package 2 will involve the creation of a hardware system designed for the automatic extraction and cleaning of earthworms from soil samples, as well as for the automatic photography of these earthworms in a straightened position.
- Work Package 3 will develop an innovative deep-learning algorithm that analyses the segmented anatomy of earthworms, treating each segment as if it were a word in a sentence, effectively considering the entire body of the earthworm as a relational combination of those segments.