JassidNet as a quantization-aware lightweight phenotyping framework for high-throughput cotton jassid (Amrasca biguttula) detection and counting toward objective resistance screening

Cotton jassid ( Amrasca biguttula ) poses a growing threat to cotton production worldwide, which creates an urgent demand for scalable and field-deployable tools that support population monitoring and resistance evaluation under natural conditions. This study proposes JassidNet, the first high-throughput computer vision-based deep learning neural network for automated cotton jassid detection and counting in real-world field environments. JassidNet integrates a lightweight detection model (SJIMNet) with standardized preprocessing, optional super-resolution enhancement, and flexible segmentation strategies, including detector-guided and SAM3-based text-driven zero-shot segmentation, to ensure robustness under heterogeneous acquisition conditions. To support model development and evaluation, a high-quality Cotton Jassid Recognition (CJR) dataset was constructed using a semi-supervised pseudo-label iteration strategy, substantially reducing manual annotation effort while maintaining label reliability. SJIMNet incorporates SPDConv, a self-customized GIAM-Block, and quantization-aware training, achieving comparable or superior detection performance relative to the YOLOv11m baseline while reducing computational cost by 20.6%, parameter count by 26.1%, and model size by 61.1%. The quantized model further compresses storage requirements to 15,388 KB while largely preserving counting-related performance, enabling efficient deployment on mobile and edge devices. The quantized model achieved 116.188 ms latency and 8.608 FPS on a CPU-only edge-computing platform, further supporting its practical deployment under resource-constrained field conditions. Explainable AI-based feature response analysis shows that, despite its lightweight design, JassidNet more effectively concentrates attention on pest-relevant regions. Extensive evaluations on both internal and independent external datasets demonstrate strong generalization capability, yielding an overall Mean Absolute Error (MAE) of 1.7, a Root Mean Square Error (RMSE) of 2.21 jassids per sampling unit, and a Pearson correlation coefficient of 0.974. Field validations further reveal a statistically consistent association between image-derived jassid counts across diverse cotton genotypes, supporting genotype-level tolerance discrimination under natural infestation pressure. Overall, JassidNet provides a practical and interpretable solution for high-throughput jassid monitoring and resistance screening, establishing a methodological foundation for intelligent plant protection and digital agriculture. Both the code and dataset can be accessed at: https://github.com/UGA-BSAIL/JassidNet .