Naslov
Cleaning Soil Invertebrate Samples using Computer Vision

Autori
Hackenberger Kutuzović, Branimir ; Đerđ, Tamara ; Hackenberger Kutuzović, Domagoj ; Hackenberger Kutuzović, Davorka

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Global soil biodiversity conference - book of abstracts / - , 2023

Skup
3rd global soil biodiversity conference

Mjesto i datum
Dublin, Irska, 13.03.2023. - 15.03.2023

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
soil mesofauna ; Berlese funnel ; deep learning ; computer vision

Sažetak
Aim: The process of quantifying and monitoring soil invertebrate biodiversity is intensive and laborious. The application of modern computer vision and deep learning techniques can accelerate the sample processing and identification processes. However, the results of computer vision-based algorithms are highly dependent on the quality of the input data. As part of this research, an algorithm was developed for the recognition and removal of soil particles from photographs of soil invertebrate samples. Method: The computer vision and deep learning-based algorithm was trained on labeled photographs of invertebrate samples containing soil particles. The convolutional neural network for soil particle segmentation was trained on a total of 100 photographs of invertebrate samples collected at different locations, with various soil types. The software was written in Python and the artificial neural network was constructed and trained using Keras. Results: Performance of the developed algorithm was evaluated using a validation dataset which consisted of input-label pairs of invertebrate samples obtained from locations with different soil types. Photographs contained in the validation dataset have not been included in the training process. The developed algorithm achieved a precision of over 85% in recognition and removal of soil particles from photographs contained in the validation dataset. The mean Intersection-over-Union of the trained algorithm was over 60% for all soil types tested. Conclusions: The developed algorithm can be used for preprocessing photographs of soil invertebrate samples, and thus improve the results of previously developed computer vision-based algorithms for automatic determination of soil invertebrate biodiversity.

Izvorni jezik
Engleski

Znanstvena područja
Biologija

POVEZANOST RADA