Integrating deep learning with digital soil core sensing for subsurface soil image segmentation

This study developed a computer vision application for soil-phase segmentation and precise characterization of soil porosity, as well as for estimating soil color and fractal properties from digital images collected in undisturbed in situ soils. The multi-sensor Digital Soil Core (DSC) was used to collect microscopic soil images from four cultivated locations in California’s Central Valley, encompassing eight soil series and six soil orders. Images were extracted from video frames of profiles down to 120 cm with 1 cm increments (1920 × 1080 px; ~3 µm/pixel; 2.2 × 1.2 mm field of view). A pre-trained (ImageNet) DeepLabV3+ model with a ResNet101 backbone was calibrated over 200 iterations using 1,564 images (90% for calibration, with an 80:20 train/validation split, and 10% for testing). Training masks were generated with a modified IsoData algorithm in ImageJ. Model performance was evaluated on an external dataset of 404 images from four profiles (0–100 cm), one per location. The final model achieved training accuracy = 0.91, loss = 0.16, and mean intersection-over-union (mIoU) = 0.86. The Precision, Recall, and F1-score on the test dataset were 0.92, 0.91, and 0.91, respectively. Model outputs (binary masks) were used to identify pore space in soil images, thereby enhancing measurements of soil porosity, fractal variables, and color analysis. We then compared three processing pipelines: unsegmented (UN), masked (MS), and binary (BI). Each pipeline produced 2D porosity (%) and fractal variables: Lacunarity, Succolarity, Fractal Dimension, and Entropy, calculated using established image-analysis algorithms. Soil color metrics (CIE L* a* b* and HSV) were computed for UN and MS images. Outputs from segmented images were compared to unsegmented images. Differences in soil color were evaluated using the paired Wilcoxon test, and differences in porosity and fractal metrics across UN, MS, and BI were assessed using the Kruskal–Wallis test. Segmentation improved accuracy and preprocessing efficiency, with statistically significant differences confirmed by these tests ( p < 0.0001). The DL-based approach is suitable for integration into digital soil sensors, enhancing segmentation accuracy and processing efficiency with inference time under 100 ms/frame. These results support scalable, field-ready quantification of soil porosity, color, and fractal metrics using in situ imagery.