Open Access

ARTICLE

AgroGeoDB-Net: A DBSCAN-Guided Augmentation and Geometric-Similarity Regularised Framework for GNSS Field–Road Classification in Precision Agriculture

Fengqi Hao^1,2,3, Yawen Hou^2,3, Conghui Gao^2,3, Jinqiang Bai^2,3, Gang Liu⁴, Hoiio Kong^1,*, Xiangjun Dong^1,2,3

1 Faculty of Data Science, City University of Macau, Macau, China
2 Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences), Jinan, China
3 Shandong Provincial Key Laboratory of Industrial Network and Information System Security, Shandong Fundamental Research Center for Computer Science, Jinan, China
4 College of Information and Electrical Engineering, China Agricultural University, Beijing, China

* Corresponding Author: Hoiio Kong. Email:

Computers, Materials & Continua 2026, 87(3), 47 https://doi.org/10.32604/cmc.2026.077252

Received 05 December 2025; Accepted 27 January 2026; Issue published 09 April 2026

Abstract

Field–road classification, a fine-grained form of agricultural machinery operation-mode identification, aims to use Global Navigation Satellite System (GNSS) trajectory data to assign each trajectory point a semantic label indicating whether the machine is performing field work or travelling on roads. Existing methods struggle with highly imbalanced class distributions, noisy measurements, and intricate spatiotemporal dependencies. This paper presents AgroGeoDB-Net, a unified framework that combines a residual BiLSTM backbone with two tightly coupled innovations: (i) a Density-Aware Local Interpolator (DALI), which balances the minority road class via density-aware interpolation while preserving road-segment structure; and (ii) a geometry-aware training objective that couples a DBSCAN-weighted focal loss with a density-regularised KL divergence, ensuring that both classification and latent representations reflect local trajectory density. The workflow first converts each enriched GNSS point into a 14-dimensional motion–spatial descriptor, projects it into a compact latent space through a variational auto-encoder, and then applies a residual BiLSTM to model bidirectional temporal dependencies before a linear classifier produces point-wise field–road predictions. Experiments on wheat, corn, and paddy datasets show overall accuracies of 98.62%, 95.46%, and 93.35%, with consistently stronger road class and overall performance than existing methods. Ablation studies further confirm that both the residual shortcut and DALI contribute positively, with DALI providing the greatest benefit for the minority road class. Tests on the unseen Harvester and Tractor datasets also demonstrate strong generalisation to previously unseen datasets. Taken together, the results show that AgroGeoDB-Net delivers reliable and scalable field–road classification from GNSS trajectories.

---

Keywords

---