SUPR
Deep learning for soil chip
Dnr:

NAISS 2024/22-226

Type:

NAISS Small Compute

Principal Investigator:

Hanbang Zou

Affiliation:

Lunds universitet

Start Date:

2024-03-01

End Date:

2025-03-01

Primary Classification:

10606: Microbiology (medical to be 30109 and agricultural to be 40302)

Webpage:

Allocation

Abstract

As the rapid elevation of carbon emission has significantly impacted our climate, one of the solutions to mitigate excessive atmospheric CO2 is soil carbon sequestration. The basic principle is that soil carbon in form of organic matter can be respired during microbial decomposition, while in other cases it can be long term stored in the soil for years. However, the conditions that govern decomposition and long-term storage are not fully understood. To investigate the mechanisms behind it, we need to look inside the soil’s pore space and investigate the conditions under which microbes are, whereas direct visualization of microorganisms succumbs to the limitation of soil opacity and the complex soil pore network. Now, microfluidic technology makes it possible. Microfluidic chips are miniaturized devices that integrate one or several analyses into a single chip, allowing real-time visualization at micro nanoscale. It is believed that the persistence of carbon in soils is affected by the organo-mineral interactions and the spatial distribution of the organic matter in the soil pore space – where microbes cannot reach their food because it may be hidden, occluded or difficult to reach. Therefore, we are making a chip that not only mimics the physical soil structure but also contain chemical heterogeneity in form of nutrient patches, such that we can interrogate the ability of microorganisms to access spaces at different levels of complexity with chemical heterogeneity. One of the major challenges for image analysis of soil chips is to recognize and extract information about the various microbes from the natural microbial community. It is previously done by manual selection, which is extremely time-consuming. Convolutional neural networks (CNNs) have been developed in the last few years to carry out tasks such as object detection, object localization, semantic segmentation, and object instance segmentation. This has led to an increased interest in the applicability of convolutional neural network-based methods for problems in medical image analysis. Recent work has shown promising results on biology and medical tasks. As such, we are applying a deep learning algorithm (Mask Rcnn/Yolvo v8) to assist our research to classify fungi, and track bacteria and protist from natural soils present as a big variety of morphologies.