A Method for Extracting Extensive Field Roads Based on Dual-Temporal Branch Net

Abstract

Field roads refer to narrow roads connecting fields with villages and between fields, facilitating agricultural activities, transportation of agricultural supplies, and farm produce. These roads are narrower and have more branches compared to regular roads, and their characteristics vary depending on the different growth stages of crops. This study considers the significant differences in characteristics of field roads at different time phases. It proposes a dual temporal branch network (Detente) for extracting field roads using two time-phase Sentinel-2 remote sensing data. A dual-branch cross-attention mechanism is designed to combine data from two time phases for road feature extraction. Additionally, dynamic snake convolution is introduced into the encoder to learn road distribution using learnable strip convolution kernels in the task of field road extraction. The effectiveness of our method was validated and analyzed on our own field road dataset and a publicly available road dataset (Deep Globe Road Dataset), with an F1-score of 0.806 and 0.799, respectively, which achieves higher accuracy compared to common road extraction models, and the method is applicable to a wide range of field road network extraction.

Keywords: Field road, Large scale road extraction, Double time phase data