@article{Qiu2026, 
author = {Zheming Qiu and Jianping Qian and Pengnan Xiao and Tianqi Lyu and Zhouqiao Ren and Zhen Wang and Chen Dong and Yiming Jiang and Boyu Shuai and Linghuan Ouyang},
title = {Method for profile horizon delineation and thickness estimation for digital soil survey},
year = {2026},
journal = {Transactions of the Chinese Society of Agricultural Engineering},
volume = {42},
number = {15},
pages = {178-188},
keywords = {soil profile segmentation, semantic segmentation, deep learning, multi-scale features, boundary localization, thickness estimation},
url = {https://www.sciopen.com/article/10.11975/j.issn.1002-6819.202512240},
doi = {10.11975/j.issn.1002-6819.202512240},
abstract = {Soil profile horizons provide important evidence for understanding pedogenic processes and environmental evolution. Accurate delineation of the surface horizon (A), subsurface horizon (B), and parent material horizon (C), together with reliable estimation of their boundary depths and thicknesses, is essential for soil classification, profile description, and digital soil survey. However, conventional horizon delineation largely depends on expert interpretation of soil color, texture, structure, root distribution, and other morphological characteristics. Such field-based judgments may vary among investigators, particularly in profiles with weak color contrasts, gradual transitions, or irregular boundaries. In addition, traditional profile descriptions are mainly recorded in textual form, and horizon thickness is usually measured manually, which limits the efficiency, consistency, and quantitative expression of soil profile information. To address these problems, this study developed a soil profile horizon segmentation and thickness estimation method based on digital images and deep learning. A total of 208 field soil profile images with clear depth scales were collected and screened from typical red soil, fluvo-aquic soil, and yellow soil profiles in Zhejiang Province, China, and were used to construct a unified dataset. Different input feature modes based on the CIE L*a*b* and HSV color spaces, as well as gray-level co-occurrence matrix texture features, were compared to improve the representation of horizon color gradients, local texture variation, and weak transitional boundaries. A semantic segmentation framework was then established to identify the A, B, and C horizons at the pixel level. On the basis of the segmentation results, the numerical marks on the profile scale were used to establish a linear mapping relationship between vertical pixel coordinates and actual soil depth. The upper and lower boundaries of each horizon were subsequently converted from pixel positions to centimeter-scale depths, and horizon thickness was calculated from the difference between adjacent boundaries. In addition, the distinctness and topography of the A-B and B-C boundaries were statistically analyzed to supplement the quantitative description of horizon boundary characteristics. The results showed that the proposed method achieved stable segmentation performance across the three soil types. The mean intersection over union was 73.18%, while the pixel accuracy was 84.64%. Among the three horizons, the A horizon showed the highest segmentation accuracy, with intersection over union values exceeding 76% for all soil types. In contrast, the segmentation accuracy of the B and C horizons decreased as boundary transitions became more gradual and structurally complex. The boundary localization results indicated that the median absolute error of the lower boundary of the A horizon was 6.45 cm, whereas the lower boundary of the B horizon had the largest error, reaching 11.26 cm. Most thickness errors for the A horizon were within 10 cm, whereas the errors for the B and C horizons were more dispersed. The analysis of boundary distinctness and transition topography further revealed systematic differences between the A–B and B-C boundaries. The A-B boundary was generally clearer and more regular, while the B-C boundary showed a higher proportion of gradual, diffuse, and irregular transitions, which was consistent with the larger localization and thickness errors observed in the lower horizons. Overall, this study established an integrated workflow from soil profile images to horizon masks, boundary depths, horizon thicknesses, and boundary structure descriptors. The method extended pixel-level segmentation results to survey-scale information and provided a practical reference for digital soil profile recording, quantitative horizon description, and assisted field soil investigation.}
}