VT-SCC: Visual-text semantic alignment via multimodal guidance for unified staple crop organ counting

Accurate counting of yield-related plant organs is fundamental to crop breeding and management. However, scalable counting remains challenging because crop morphology varies widely and field environments are complex. Current specialized models lack flexibility, whereas generic few-shot counters often exhibit feature ambiguity arising from their reliance on visual exemplars, leading to frequent misidentification in dense scenes. To address these challenges, we developed VT-SCC, a visual-text unified staple crop organ counting framework that incorporates semantic guidance into visual representation learning. VT-SCC integrates a Scale-Adaptive Feature Aggregator (SAFA), which uses a multi-branch structure to adaptively capture targets with large size variation, and a Semantic-Guided Channel Modulator (SGCM), which uses category text prompts to calibrate visual features dynamically and suppress background interference. In extensive experiments on a diverse dataset comprising wheat (Triticum aestivum), maize (Zea mays), rice (Oryza sativa), and soybean (Glycine max), VT-SCC achieved a mean absolute error(MAE) of 3.28 and a root mean square error(RMSE) of 4.48, outperforming existing generic counters while remaining competitive with specialized models. VT-SCC also generalized effectively in few-shot transfer learning tasks, adapting to unseen domains with minimal data. These observations demonstrate the feasibility of VT-SCC as a unified agricultural counter and highlight its potential as a scalable, robust solution for high-throughput phenotyping. The code and dataset are publicly available at http://vt-scc.samlab.cn.