MS-TFNet: a multi-scale temporal feature fusion network for maize growth stage recognition

Abstract

Accurate maize growth stage recognition is critical for precision agricultural management, but it remains challenging due to subtle morphological differences between adjacent stages, small seedling targets, and substantial cross-year environmental variability. To address these challenges, this study proposes a Multi-Scale Temporal Feature Fusion Network, termed MS- TFNet, for image-level maize growth stage recognition using red-green-blue (RGB) images and acquisition-date information. The proposed network integrates three complementary components. First, a Multi-Scale Feature Extraction Module (MS) is introduced to capture maize structural variations at different receptive-field scales. Second, a Vegetation-Aware Convolutional Block Attention Module (VCBAM) is designed to adaptively refine RGB feature responses and improve the representation of maize-related visual cues under complex field backgrounds. Third, a Temporal Feature Module (TFM) encodes image acquisition dates using date-based sine–cosine temporal encoding, enabling phenological prior information to be incorporated into single-image growth stage classification. Experiments were conducted on a long-term field observation dataset containing 24,239 maize images collected over eight growing seasons and covering nine maize growth stages. The results show that MS-TFNet achieved a mean Overall Accuracy (OA) of 83.64%, outperforming the ResNet-50 baseline by 4.46 percentage points. Under a ±3-day temporal tolerance criterion, the recognition accuracy further increased to 93.76%, indicating improved robustness for practical crop monitoring scenarios. Overall, MS-TFNet provides an effective framework for maize growth stage recognition under complex field conditions by jointly exploiting multi-scale visual representation, vegetation-aware attention, and acquisition-date-based phenological priors.