Use of deep learning for structural analysis of computer tomography images of soil samples.

Soil samples from several European countries were scanned using medical computer tomography (CT) device and are now available as CT images. The analysis of these samples was carried out using deep learning methods. For this purpose, a VGG16 network was trained with the CT images (X). For the annotation (y) a new method for automated annotation, 'surrogate' learning, was introduced. The generated neural networks (NNs) were subjected to a detailed analysis. Among other things, transfer learning was used to check whether the NN can also be trained to other y-values. Visually, the NN was verified using a gradient-based class activation mapping (grad-CAM) algorithm. These analyses showed that the NN was able to generalize, i.e. to capture the spatial structure of the soil sample. Possible applications of the models are discussed.

Proximal Soil Sensing - A Contribution for Species Habitat Distribution Modelling of Earthworms in Agricultural Soils?

BACKGROUND: Earthworms are important for maintaining soil ecosystem functioning and serve as indicators of soil fertility. However, detection of earthworms is time-consuming, which hinders the assessment of earthworm abundances with high sampling density over entire fields. Recent developments of mobile terrestrial sensor platforms for proximal soil sensing (PSS) provided new tools for collecting dense spatial information of soils using various sensing principles. Yet, the potential of PSS for assessing earthworm habitats is largely unexplored. This study investigates whether PSS data contribute to the spatial prediction of earthworm abundances in species distribution models of agricultural soils. METHODOLOGY/PRINCIPAL FINDINGS: Proximal soil sensing data, e.g., soil electrical conductivity (EC), pH, and near infrared absorbance (NIR), were collected in real-time in a field with two management strategies (reduced tillage / conventional tillage) and sandy to loam soils. PSS was related to observations from a long-term (11 years) earthworm observation study conducted at 42 plots. Earthworms were sampled from 0.5 x 0.5 x 0.2 m³ soil blocks and identified to species level. Sensor data were highly correlated with earthworm abundances observed in reduced tillage but less correlated with earthworm abundances observed in conventional tillage. This may indicate that management influences the sensor-earthworm relationship. Generalized additive models and state-space models showed that modelling based on data fusion from EC, pH, and NIR sensors produced better results than modelling without sensor data or data from just a single sensor. Regarding the individual earthworm species, particular sensor combinations were more appropriate than others due to the different habitat requirements of the earthworms. Earthworm species with soil-specific habitat preferences were spatially predicted with higher accuracy by PSS than more ubiquitous species. CONCLUSIONS/SIGNIFICANCE: Our findings suggest that PSS contributes to the spatial modelling of earthworm abundances at field scale and that it will support species distribution modelling in the attempt to understand the soil-earthworm relationships in agroecosystems.