A novel deep learning approach to field-road semantic segmentation.

The automatic segmentation of field-road using artificial intelligence (AI) is imperative for intelligence agriculture, allowing for the distinction between operational patterns (e.g., turning and transporting) through the analysis of global navigation satellite system (GNSS) data. This AI-driven discrimination is crucial for accurately monitoring the field operations of agricultural machinery. This study presents a deep learning framework that implements a novel semantic segmentation model designed to segment field-road trajectories, addressing the marked differences in spatial characteristics. The developed AI approach integrates transformer and semantic technologies to create an advanced semantic encoder that generates high-quality semantic prior maps and associated mask features. These features are effectively combined through a novel lightweight up-sampling mechanism paired with a semantic feature pyramid network (FPN) decoder, resulting in improved prediction outputs. The class imbalance issue between field-road pixels is effectively addressed by employing a pixel-wise weighted cross-entropy loss function in this study. Additionally, the model identifies unique features by evaluating GNSS points' similarities to adjacent points and their global class counterparts. The proposed method was evaluated using the dataset comprising 6,380 GNSS trajectory images of wheat and rice in this study. The experimental results demonstrate that the mean intersection-over-union (mIoU) and F1-score of the model achieved 92.46% and 92.65%, respectively. Consequently, this study contributes significantly to refined field-operation cost analysis and is instrumental for advancements in precision mechanization management and agricultural intelligence.