Bao et al. (2026) An edge-effect-aware framework for uncertainty-resolved crop mapping and planting-structure monitoring in arid irrigation landscapes

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-10-17
• Authors: Lu Bao, Sinan Wang, Ruiping Li, Yingjie Wu, Wenjuan Wang, Xianlin Chen, Jian Liu, Xiaoming Ma
• DOI: 10.1016/j.agwat.2026.110850

Research Groups

• Institute of Pastoral Hydraulic Research, MWR, Hohhot, China
• Yinshanbeilu Grassland Eco-Hydrology National Observation and Research Station, China Institute of Water Resources and Hydropower Research, Beijing, China
• College of Water Conservancy and Civil Engineering, Inner Mongolia Agricultural University, Hohhot, Inner Mongolia, China
• State Key Laboratory of Water Engineering Ecology and Environment in Arid Area, Inner Mongolia Agricultural University, Hohhot, Inner Mongolia, China
• Inner Mongolia JinHua Yuan Environmental Resource Engineering Consulting Co., Ltd, Hohhot, Inner Mongolia, China
• School of Civil and Hydraulic Engineering, Hefei University of Technology, Hefei, China
• Water Resources Research Institute of Inner Mongolia Autonomous Region, Hohhot, Inner Mongolia, China

Short Summary

This study developed an edge-effect-aware framework for uncertainty-resolved crop mapping and planting-structure monitoring in arid irrigation landscapes. It found that a structurally simple Standard Neural Network (SNN) combined with spatial filtering achieved superior accuracy (0.953 Overall Accuracy) and robustness against mixed-pixel interference, revealing a "sunflower matrix and maize aggregation" pattern driven by economic incentives and water constraints.

Objective

• To systematically evaluate and compare the categorization accuracy of classical machine learning (Random Forest) and deep learning models (Standard Neural Network, One-Dimensional Convolutional Neural Network, and Long Short-Term Memory) based on a unified Optuna optimization framework, and clarify the role of spatial filtering in achieving high-fidelity plot reconstruction.
• To analyze the fine-scale spatiotemporal evolution patterns of major crop planting structures in the Yichang Irrigation Area from 2023 to 2025, revealing the dynamic adjustments under water and soil resource constraints.
• To quantify the constraint mechanisms of edge effects on classification accuracy in fragmented landscapes, resolving the differences and uncertainty sources of different models in countering mixed pixel interference from the dimensions of boundary distance and spatiotemporal consistency.

Study Configuration

• Spatial Scale: Yichang Irrigation Area (40°41′–41°11′N and 107°25′–108°58′E), a sub-region of the Hetao Irrigation District, Yellow River Basin, China. Pixel resolution of 10 meters for Sentinel-2 imagery.
• Temporal Scale: Crop growing seasons from July 1st to September 30th for the years 2023, 2024, and 2025 for classification. Auxiliary time-series data from May to October for phenological analysis.

Methodology and Data

• Models used: Standard Neural Network (SNN, equivalent to Multi-Layer Perceptron), One-Dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM) network, Random Forest (RF). Optuna Bayesian optimization framework for hyperparameter tuning. Spatial Majority Voting Filter for post-processing.
• Data sources:
  • Sentinel-2 Multi-Spectral Instrument (MSI) Level-2A surface reflectance products (ESA Copernicus Open Access Hub).
  • Derived spectral indices: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Difference Water Index (NDWI).
  • Ground truth data: Field surveys conducted synchronously during peak reproductive stage (July–September) from 2023 to 2025, with 7 land cover classes (sunflower, maize, wheat, other crops, bare land, built-up land, and water).
  • Auxiliary Time-Series Data: Monthly NDVI from May to October for 2023–2025 for phenological analysis.

Main Results

• The structurally simple SNN model, after Optuna optimization and a 3x3 spatial majority voting filter, achieved the highest Overall Accuracy (OA) of 0.953 (95% CI: 0.953 ± 0.006), outperforming 1D-CNN (0.945), RF (0.936), and LSTM (0.926).
• The SNN demonstrated superior robustness in processing static spectral-index features, avoiding overfitting issues encountered by LSTM with non-temporal mixed spectral signals in fragmented landscapes.
• Spatiotemporal analysis (2023–2025) revealed a "sunflower matrix and maize aggregation" pattern. Maize area continuously expanded from 78,400 hectares in 2023 to 98,600 hectares in 2025, while sunflower area declined from 230,100 hectares to 211,600 hectares, driven by economic incentives and water constraints.
• Boundary-distance analysis confirmed that mixed pixels within the 0–10 meter range from field boundaries constitute the primary physical bottleneck limiting classification accuracy (OA < 0.80). SNN exhibited significantly better stability in these transition zones (OA ≈ 0.78) compared to LSTM (OA ≈ 0.73).
• Derived spectral indices (NDVI, NDWI, EVI) were identified as the most critical features for crop discrimination, outweighing raw spectral bands.

Contributions

• Developed an edge-effect-aware framework that integrates Sentinel-2 data, Optuna-optimized deep learning models, and spatial filtering for uncertainty-resolved crop mapping in fragmented arid irrigation landscapes.
• Demonstrated that a structurally simple SNN can outperform more complex recurrent networks (LSTM) for static feature-based crop classification in highly fragmented areas with significant mixed-pixel issues, providing new insights into model selection for such environments.
• Quantified the impact of "edge effects" on classification accuracy, highlighting the 0–10 meter boundary zone as a critical bottleneck and emphasizing the need for uncertainty-resolved mapping beyond global accuracy metrics.
• Provided fine-scale spatiotemporal monitoring of crop planting structures (2023–2025) in the Yichang Irrigation Area, revealing dynamic adjustments driven by economic incentives and water-soil resource constraints, offering support for adaptive planting-structure regulation.

Funding

• Inner Mongolia Autonomous Region Science and Technology Plan Project (2023YFHH0109)
• Key Project of the National Natural Science Foundation of China - Yellow River Water Science Research Joint Fund (U2443210)
• The Special Training Program Project for Outstanding Young Talents under the “Talents Revitalizing Inner Mongolia” Project of the Inner Mongolia Autonomous Region The 2022 Scientific Research Support Project for the Introduction of High-level Talents at the Inner Mongolia Autonomous Region Level.

Citation

@article{Bao2026edgeeffectaware,
  author = {Bao, Lu and Wang, Sinan and Li, Ruiping and Wu, Yingjie and Wang, Wenjuan and Chen, Xianlin and Liu, Jian and Ma, Xiaoming},
  title = {An edge-effect-aware framework for uncertainty-resolved crop mapping and planting-structure monitoring in arid irrigation landscapes},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110850},
  url = {https://doi.org/10.1016/j.agwat.2026.110850}
}