A smartphone-based approach for comprehensive soil microbiome profiling

Abstract

The soil microbiome is crucial for nutrient cycling, health, and plant growth. This study presents a smartphone-based approach as a low-cost and portable alternative to traditional methods for classifying bacterial species and characterizing microbial communities in soil samples. By harnessing bacterial autofluorescence detection and machine learning algorithms, the platform achieved an average accuracy of 88% in distinguishing common soil-related bacterial species despite the lack of biomarkers, nucleic acid amplification, or gene sequencing. Furthermore, it successfully identified dominant species within various bacterial mixtures with an accuracy of 76% and three-level soil health identification at an accuracy of 80%-82%, providing insights into microbial community dynamics. The influence of other soil conditions (pH and moisture) was relatively minor, showcasing the platform's robustness. Various field soil samples were also tested with this platform at 80% accuracy compared with the laboratory analyses, demonstrating the practicality and usability of this approach for on-site soil analysis. This study highlights the potential of the smartphone-based system as a valuable tool for soil assessment, microbial monitoring, and environmental management.

FIG. 1.

The overall procedure of smartphone-based profiling of soil bacteria. A soil sample is collected, dissolved in DI water, and centrifuged. The supernatant is added to a glass slide using a transfer pipette, and the smartphone-based imaging system with an acrylic filter wheel (12 filters) and three LEDs captures multiple images of the sample. Each image is split into nine sub-parts, and the RGB values are extracted from the sample areas. The collected RGB intensities are used to build the learning database, and the prediction model is developed using various machine learning algorithms (decision tree, k-nearest neighbors, support vector machines, and extreme gradient boosting). All data were randomly split to 80% training and 20% test. Initially, the suspension of individual bacterial species is tested, followed by testing 40 or 51 combinations of various bacterial mixtures spiked into the soil. Finally, field soil samples are tested. A dominant species, three- to five-level soil health, or presence of pathogenic bacteria is identified from the machine learning model.

FIG. 2.

Device schematic and its operation. (a) An LED box excites the bacterial solution collected from the soil on a glass slide, and a smartphone equipped with an acrylic filter wheel captures autofluorescence images. (b) 12 different raw smartphone images are collected from a single soil sample. (c) A single image is split into nine parts. (d) RGB values are extracted from each part of an image. (e) All RGB information is collected to build the database for machine learning classification. A confusion matrix shows an accuracy of 88.33% in determining bacterial species (cultured in LB to 10⁸ CFU/ml) (80%/20% random splits of training and test sets). (f) Prediction results are shown by each bacterial species.

FIG. 3.

Identification of dominant species from bacterial mixtures in the soil and primarily contributed features. (a) Confusion matrix and prediction histograms by each bacterial species for predicting dominant species. The accuracy is 76.02%. (b) Important contributions of top 20 features in classifying four bacterial species. (c) Heatmap summary of top ten features contributed most to the classifications, using SHAP summary function to get the average absolute SHAP values (numpy.arrry; left) or the global importance of each feature is taken to be the mean absolute SHAP values (shap.Explainer; right) over all samples for all features. (d) List of all 36 features with varying acrylic filter types, LED light sources, and the color channels (R, G, or B) of smartphone images. (e) Five major contributed features (filter and LED) for bacterial classification and their possible intrinsic fluorophores.

FIG. 4.

Soil health classification from the simulated soil samples. Three levels (A, B, and C) were set based on the composition ratio between beneficial and pathogenic bacteria cultured in soil samples. (a) and (b) Compositions of binary mixtures (one beneficial and one pathogenic) and XGBoost classification result. (c) and (d) Compositions of four-species mixtures (two beneficial and two pathogenic) and XGBoost classification results. The accuracies were calculated using independent validation sets.

FIG. 5.

Soil health classification under different pH and moisture conditions. (a) Confusion matrices. The first letter represents pH: A = acidic, N = neutral, and B = basic. The second letter represents moisture content: L = unavailable, M = plant-available, and H = oversaturated. (b) Bar plots of classification accuracies under nine conditions.

FIG. 6.

Tests with field soil samples. (a) Pin-styled map showing the sample collection spots. Blue indicates that the sample contains E. coli, while orange indicates that the sample contains S. enterica. Gray pins indicate neither E. coli nor S. enterica. (b) pH and moisture values were tested using a soil meter. The E. coli presence was identified with an ELISA kit, and the S. enterica presence was confirmed with selective agar plates. (c) Prediction results with a smartphone-based system and machine learning classification. Incorrect predictions from dominant species classification (compared with the ELISA and selective agar plate results) are labeled as “x”: samples #3, #11, #13, #17, #25, and #27, i.e., 80% accuracy. Soil health levels are also predicted: dark green = level A (pathogenic), medium green = level B (medium), and light green = level C (beneficial).