Laplacian Frequency Interaction Network for Rural Thematic Road Extraction

Problem
This paper addresses the challenge of extracting topological road structures from movement trajectory images of agricultural machinery, a task critical for rural thematic road network construction. Existing methods often rely on downsampling techniques that blur high-frequency road structures, leading to fragmented or redundant topologies due to noise from dense field operations. The authors propose a novel approach, LFINet, to overcome these limitations. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution is the Laplacian Frequency Interaction Network (LFINet), which employs a dual-pathway architecture to effectively separate and process low-frequency semantic contexts and high-frequency structural details. The architecture consists of the following components:

1. Laplacian Multi-scale Separator (LMS): This module decouples the input image into low-frequency and high-frequency components.
2. Cross-Frequency Interaction Block (CFIB): This block features two pathways:
   • High-Frequency Block (HFB): Refines local structural details.
   • Spatial Transformer (ST): Captures global semantic information.
3. Frequency Gated Modulation (FGM): This mechanism integrates features from both pathways, using semantic contexts to modulate structural details.
4. Progressive Reconstruction Decoder: This component iteratively fuses multi-scale features to maintain topological consistency in the final output.

The training process and compute resources utilized are not explicitly disclosed in the paper.

Results
LFINet achieves state-of-the-art performance on a real-world dataset of agricultural trajectories from Henan Province, China. The model reports an F1-score of 92.54% and an Intersection over Union (IoU) of 86.12%. These results surpass the second-best method by 0.64% in F1-score and 1.1% in IoU, demonstrating the effectiveness of LFINet in constructing topological road networks from noisy and sparse data.

Limitations
The authors acknowledge that the performance of LFINet may be influenced by the quality and diversity of the training dataset, which is limited to a specific geographical region (Henan Province). They do not discuss potential generalization issues to other rural contexts or the computational efficiency of the model in real-time applications. Additionally, the reliance on a dual-pathway architecture may increase model complexity and training time, which could be a concern for deployment in resource-constrained environments.

Why it matters
The development of LFINet has significant implications for the field of remote sensing and agricultural monitoring. By effectively extracting road networks from noisy trajectory data, this work can enhance the accuracy of rural infrastructure mapping and support better decision-making in agricultural logistics and planning. Furthermore, the methodologies introduced, particularly the Laplacian frequency separation and cross-frequency interaction, may inspire future research in other domains requiring high-fidelity feature extraction from complex data.

Authors: Baiyan Chen, Weixin Zhai
Source: arXiv:2605.02866
URL: https://arxiv.org/abs/2605.02866v1