Zeng et al. (2025) Stratified layering for soil profile: Dynamic short field Mamba network

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2025
• Date: 2025-11-13
• Authors: Shaohua Zeng, Zhihao Chen, Ruolan Zeng, Shuai Wang, Yang Wang
• DOI: 10.1016/j.compag.2025.111212

Research Groups

• College of Computer & Information Science, Chongqing Normal University, China
• Chongqing Centre of Engineering Technology Research On Digital Agricultural & Service, China
• Chongqing Electric Power College, China
• Chongqing Master Station of Agricultural Technology Promotion, China
• Jiangjin District Center On Agricultural Technology Promotion, Jiangjin, Chongqing, China

Short Summary

This paper proposes a Dynamic Short Field Mamba (DSFM) network, built upon the Vision Mamba UNet, for accurate stratification of soil profiles, achieving significant improvements in accuracy by optimizing spatial feature extraction and handling blurred inter-layer transitions.

Objective

• To develop an accurate and automated computer vision method for stratifying soil profiles, specifically addressing the challenges posed by multi-layered structures and indistinct boundaries in soil profile images.

Study Configuration

• Spatial Scale: Image-level analysis of soil profiles, focusing on local regions and adjacent layers within an image.
• Temporal Scale: Not explicitly mentioned as a variable in the study; the model processes static images.

Methodology and Data

• Models used:
  • Dynamic Short Field Mamba (DSFM) network (proposed)
  • Vision Mamba UNet (VM-UNet) (baseline architecture)
  • Similarity Encoder of Short Field (SESF) module (component of DSFM)
  • Dynamic Position Encoder (DPE) module (component of DSFM)
• Data sources:
  • Soil profile images (specific dataset details not provided in the excerpt).

Main Results

• The DSFM network achieved an Overall Accuracy (OA) of 78.99 % and a mean Intersection over Union (mIoU) of 60.22 % on the soil profile stratification task.
• These results represent an improvement of 6.98 % in OA and 10.46 % in mIoU compared to the original VM-UNet.
• The Similarity Encoder of Short Field (SESF) module effectively optimizes spatial feature extraction by computing similarity only between adjacent soil layers, thus avoiding redundant cross-layer computations.
• The Dynamic Position Encoder (DPE) module successfully handles blurred and gradual transitions between soil layers by integrating short field similarity and static positional information to construct a dynamic position matrix.

Contributions

• Proposes the Dynamic Short Field Mamba (DSFM) network, a novel architecture specifically designed for accurate and automated soil profile stratification.
• Introduces the Similarity Encoder of Short Field (SESF) module to enhance spatial feature extraction for multi-layered soil profiles by focusing on local, adjacent layer similarities.
• Develops the Dynamic Position Encoder (DPE) module to effectively manage blurred and gradual transitions between soil layers, improving the model's sensitivity to inter-layer positional variations.
• Demonstrates a significant improvement in stratification accuracy (6.98 % OA, 10.46 % mIoU) over the baseline VM-UNet, addressing a critical gap in automated soil profile analysis.

Funding

• Not specified in the provided text.

Citation

@article{Zeng2025Stratified,
  author = {Zeng, Shaohua and Chen, Zhihao and Zeng, Ruolan and Wang, Shuai and Wang, Yang},
  title = {Stratified layering for soil profile: Dynamic short field Mamba network},
  journal = {Computers and Electronics in Agriculture},
  year = {2025},
  doi = {10.1016/j.compag.2025.111212},
  url = {https://doi.org/10.1016/j.compag.2025.111212}
}