Sheng Xu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sheng Xu has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Chunling Chen
College of Information and Electronic Engineering, Shenyang Agricultural University, Shenyang 110866, China
1 shared publication - Shuai Feng
College of Information and Electrical Engineering, Shenyang Agricultural University, Shenyang 110866, China
1 shared publication - Tongyu Xu
College of Information and Electrical Engineering, Shenyang Agricultural University, Shenyang 110866, China
1 shared publication - Yongsheng Liu
College of Information and Electrical Engineering, Shenyang Agricultural University, Shenyang 110866, China
1 shared publication - Zhonghui Guo
College of Information and Electrical Engineering, Shenyang Agricultural University, Shenyang 110866, China
1 shared publication
Latest Publications
- Open Access
- 16 September 2026
- Published by MDPI AG in Agriculture
- Vol. 16 (18) , 1986
An Improved DeepLabv3+-Based Framework for Field-Road Extraction and Structural Indicator Quantification in Well-Facilitated Farmland
Well-facilitated farmland plays an important role in stabilizing and increasing crop yields, advancing agricultural mechanization, and improving production efficiency. Field-road quality directly affects machinery access and the transport of agricultural inputs and harvested crops. Current acceptance inspections rely mainly on field surveys and spot measurements, resulting in limited spatial coverage, low efficiency, and poor reproducibility. Therefore, a method is needed to rapidly survey entire field-road networks, extract road extents and structural indicators, and generate verifiable inspection records. Centimeter-resolution UAV imagery enables flexible and repeatable data acquisition, but deriving acceptance-oriented indicators remains challenging. Narrow field roads are readily obscured by crops and shelterbelt shadows or confused with cropland textures, resulting in blurred boundaries, discontinuities, and false detections. We developed a lightweight UAV-based framework integrating field-road segmentation and structural-indicator quantification. MobileNetV2 replaced the DeepLabv3+ backbone, while a Normalization-based Attention Module and Content-Aware ReAssembly of FEatures enhanced interference suppression and spatial reconstruction. Morphological processing, skeleton extraction, Euclidean distance transformation, and skeleton-graph analysis were then used to quantify road width and network connectivity. Across three training runs with different random seeds, the model achieved mean mIoU, mPA, and precision values of 93.34%, 96.75%, and 98.90%, respectively. The model had 6.14 million parameters and an inference speed of 17.04 FPS. After averaging five measurements from each road segment, the R2 values between the predicted and manually measured widths were 0.650, 0.486, and 0.662 for asphalt, concrete, and gravel roads, respectively. The corresponding width MAEs were 0.130, 0.140, and 0.100 m. Connectivity analysis yielded an index of 1.00 in the first validation area, while gap repair increased the index from 0.4682 to 0.4795 in the second area. The framework supports efficient, quantitative, and traceable acceptance inspection of field-road infrastructure in well-facilitated farmland.
- 16 September 2026
- Published by MDPI AG in Agriculture
- Vol. 16 (18) , 1986
Open AccessAn Improved DeepLabv3+-Based Framework for Field-Road Extraction and Structural Indicator Quantification in Well-Facilitated Farmland
Well-facilitated farmland plays an important role in stabilizing and increasing crop yields, advancing agricultural mechanization, and improving production efficiency. Field-road quality directly affects machinery access and the transport of agricultural inputs and harvested crops. Current acceptance inspections rely mainly on field surveys and spot measurements, resulting in limited spatial coverage, low efficiency, and poor reproducibility. Therefore, a method is needed to rapidly survey entire field-road networks, extract road extents and structural indicators, and generate verifiable inspection records. Centimeter-resolution UAV imagery enables flexible and repeatable data acquisition, but deriving acceptance-oriented indicators remains challenging. Narrow field roads are readily obscured by crops and shelterbelt shadows or confused with cropland textures, resulting in blurred boundaries, discontinuities, and false detections. We developed a lightweight UAV-based framework integrating field-road segmentation and structural-indicator quantification. MobileNetV2 replaced the DeepLabv3+ backbone, while a Normalization-based Attention Module and Content-Aware ReAssembly of FEatures enhanced interference suppression and spatial reconstruction. Morphological processing, skeleton extraction, Euclidean distance transformation, and skeleton-graph analysis were then used to quantify road width and network connectivity. Across three training runs with different random seeds, the model achieved mean mIoU, mPA, and precision values of 93.34%, 96.75%, and 98.90%, respectively. The model had 6.14 million parameters and an inference speed of 17.04 FPS. After averaging five measurements from each road segment, the R2 values between the predicted and manually measured widths were 0.650, 0.486, and 0.662 for asphalt, concrete, and gravel roads, respectively. The corresponding width MAEs were 0.130, 0.140, and 0.100 m. Connectivity analysis yielded an index of 1.00 in the first validation area, while gap repair increased the index from 0.4682 to 0.4795 in the second area. The framework supports efficient, quantitative, and traceable acceptance inspection of field-road infrastructure in well-facilitated farmland.