Research

Similar symptoms can occur on different crops, while clutter and occlusion make field recognition difficult. TaxoSeg combines field images with a crop–disease taxonomy to locate lesions at pixel level and identify their crop and disease category.

Structured semantic prototypes connect visual regions to category knowledge. Crop context helps distinguish similar symptoms, while class names and symptom descriptions jointly guide region recognition.

We study quantized low-rank adaptation of the official Qwen3.5-0.8B Instruct multimodal base for agriculture. Agricultural images and category supervision support our investigation of crop and crop-health recognition in compact models.

The model reads an image and outputs one Chinese category name. This explicit output provides an entry point for organizing agricultural knowledge and retrieving relevant literature by category.

A vision–language model first identifies a Chinese category name from the image. That category and the user’s question are then used to retrieve agricultural literature. The response is organized around the retrieved evidence.

When lesion extent is needed, TaxoSeg supplies segmentation results. Category recognition, lesion locations and literature evidence together support agricultural analysis. The overall pipeline is currently being integrated as a research project.

Post-translational modifications are closely connected to protein function. PlantPTM integrates protein language models with multi-view features to predict diverse modification sites from plant protein sequences, providing a computational tool for plant protein research.

The work is published in Molecular Plant. Source code and an online prediction platform are available for use and further research.