Zhang et al. (2025) Comparative Analysis of Deep Learning and Traditional Methods for High-Resolution Cropland Extraction with Different Training Data Characteristics

Identification

• Journal: Land
• Year: 2025
• Date: 2025-10-13
• Authors: Dujuan Zhang, Xiufang Zhu, Yaozhong Pan, Hengliang Guo, Qiannan Li, Haitao Wei
• DOI: 10.3390/land14102038

Research Groups

• National Supercomputing Center in Zhengzhou, Zhengzhou University, Zhengzhou, China
• Institute of Remote Sensing Science and Engineering, Faculty of Geographical Sciences, Beijing Normal University, Beijing, China
• School of Water Conservancy and Transportation, Zhengzhou University, Zhengzhou, China
• State Key Laboratory of Remote Sensing Science, Beijing Normal University, Beijing, China
• The 27th Research Institute of China Electronics Technology Group Corporation, Zhengzhou, China
• Henan Key Laboratory of Remote Sensing of Ecological Environment, Zhengzhou, China
• School of Geoscience and Technology, Zhengzhou University, Zhengzhou, China

Short Summary

This study comparatively analyzes deep learning (UNet, DeepLabv3+) and traditional (OBIA-RF) methods for high-resolution cropland extraction, evaluating the impact of classifier choice, band combinations, crop growth stages, and training data mislabeling. Deep learning models consistently outperformed traditional methods, demonstrating higher robustness to varying data characteristics and complex landscapes.

Objective

• How do the performances of two common DCNN architectures, UNet and DeepLabv3+, compare with each other and with the conventional OBIA-RF method in HRRS cropland mapping?
• How do different band combinations impact the accuracy of cropland extraction?
• How do HRRS images from various crop growth stages affect the performance of cropland mapping?
• What are the implications of class mislabeling on the accuracy of cropland classification?

Study Configuration

• Spatial Scale: Two county-wide regions in Shandong Province, China: Juye County (1302 km²) and Qixia County (1793 km²). High-resolution remote sensing (HRRS) imagery with 2 m spatial resolution.
• Temporal Scale: Gaofen-1 satellite images acquired at different crop growth stages: non-vigorous (Juye: 20 October 2015; Qixia: 25 June 2016) and vigorous (Juye: 1 April 2016; Qixia: 22 August 2016).

Methodology and Data

• Models used:
  • Deep Learning: UNet (ResNet-50 backbone), DeepLabv3+ (ResNet-50 backbone, Atrous Spatial Pyramid Pooling - ASPP module).
  • Traditional: Object-Based Image Analysis with Random Forest (OBIA-RF).
• Data sources:
  • Satellite imagery: Gaofen-1 (GF-1) PMS (Panchromatic and Multispectral) cameras.
  • Image processing: 2 m spatial resolution composite images (fusion of 2 m PAN and 8 m MS using Gram Schmidt transformation).
  • Spectral bands: Panchromatic (450–900 nm), Blue (450–520 nm), Green (520–590 nm), Red (630–690 nm), Infrared (770–890 nm).
  • Reference data: Cropland sample datasets annotated through visual interpretation of GF-1 composite data by professionals, cross-validated, and supplemented with Google Maps.

Main Results

• UNet and DeepLabv3+ models consistently outperformed OBIA-RF for HRRS cropland extraction across various crop growth stages in both simple (plain) and complex (mountainous) agricultural landscapes. In complex mountainous terrain, DL models showed substantial improvements in cropland's Producer's Accuracy (PA) (up to 38.31% for UNet, 33.95% for DeepLabv3+).
• Deep learning models (UNet, DeepLabv3+) were largely insensitive to different band combinations (Near-Infrared–Red–Green–Blue, Near-Infrared–Red–Green, and Red–Green–Blue), indicating their ability to learn abstract features. OBIA-RF's performance varied, with Near-Infrared (NIR) information being crucial for its accuracy.
• Cropland extraction accuracies varied across different crop growth stages, with smaller discrepancies in plain areas compared to mountainous regions. OBIA-RF was more sensitive to temporal changes than UNet and DeepLabv3+.
• All three models (UNet, DeepLabv3+, OBIA-RF) remained relatively resilient to training data noise with up to 5% mislabeling. Beyond this threshold, classification accuracy declined, with OBIA-RF experiencing a more significant drop than UNet and DeepLabv3+.
• Data augmentation strategies (flipping, rotation) slightly improved HRRS cropland extraction accuracies in plain areas but marginally reduced them in complex mountainous areas.
• UNet and DeepLabv3+ exhibited relatively low sensitivity to the patch size of sample blocks (beyond a certain threshold, e.g., 256 × 256 pixels) and the depth of the ResNet backbone architecture (ResNet18 to ResNet152).

Contributions

• Provides valuable insights into the impact of classifier selection, band combinations, temporal characteristics, and class mislabeling on high-resolution cropland extraction across diverse agricultural landscapes (plain and mountainous regions).
• Offers guidance for researchers in designing experiments for high-resolution cropland extraction.
• Facilitates informed decision-making for agricultural practitioners regarding efficient, timely, and accurate cropland identification.
• Supports policymakers with reliable cropland monitoring data for developing effective agricultural policies and strategies that support food security and sustainable development goals.
• Demonstrates the superior robustness and stability of deep learning algorithms (UNet, DeepLabv3+) compared to traditional methods (OBIA-RF) in the presence of noisy labels and varying data characteristics.

Funding

• National Natural Science Foundation of China (No. 42501427)
• Key Research and Development Special Projects in Henan Province (No. 241111212300, No. 221111320600)
• Key R&D and Promotion Special Projects of Henan Province (No. 242102321024)
• Postgraduate Education Reform and Quality Improvement Project of Henan Province (No. YJS2025GZZ06)

Citation

@article{Zhang2025Comparative,
  author = {Zhang, Dujuan and Zhu, Xiufang and Pan, Yaozhong and Guo, Hengliang and Li, Qiannan and Wei, Haitao},
  title = {Comparative Analysis of Deep Learning and Traditional Methods for High-Resolution Cropland Extraction with Different Training Data Characteristics},
  journal = {Land},
  year = {2025},
  doi = {10.3390/land14102038},
  url = {https://doi.org/10.3390/land14102038}
}