Bridging convolution and attention: An ultra-lightweight hybrid model for freeze damage classification in camelina

Authors: NIJHUM, PAUL - North Dakota State University; SUNIL, GC - North Dakota State University; ALI, AIZAZ - North Dakota State University; Anderson, James; Sthapit Kandel, Jinita; RAHMAN, MUKHLESUR - North Dakota State University; Horvath, David; SUN, XIN - North Dakota State University

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/28/2025
Publication Date: 8/29/2025
Citation: Nijhum, P., Sunil, G., Ali, A., Anderson, J.V., Sthapit Kandel, J., Rahman, M., Horvath, D.P., Sun, X. 2025. Bridging convolution and attention: An ultra-lightweight hybrid model for freeze damage classification in camelina. Smart Agricultural Technology. https://doi.org/10.1016/j.atech.2025.101396.
DOI: https://doi.org/10.1016/j.atech.2025.101396

Interpretive Summary: Quantifying the amount of damage caused by freezing in camelina is difficult, slow, and subjective. To overcome this problem, a series of deep learning models were trained and tested on a diverse set of images of camelina plants that had variable damage levels following different preliminary growth conditions and freezing treatments. Various models were tasked with identifying the differences between the plant leaves and the background, and then were trained ton 3,114 manually annotated images to classify individual pixels on the leaves as either dead, damaged, or undamaged. A custom model proved to be the best for this task. The results of this work will allow the development of a simple application to quantify freezing damage in camelina using high-throughput methods with limited biases.

Technical Abstract: Freeze damage significantly threatens camelina crop productivity, necessitating rapid, accurate, and scalable diagnostic methods. Traditional approaches to assessing freeze injury are labor-intensive, subjective, and impractical for large-scale implementation. This study proposes a novel deep learning- based model that integrates a convolutional backbone with a transformer-based attention mechanism to automate freeze damage classification from RGB images. The model leverages the local feature extraction capabilities of CNNs and the global context modeling strengths of transformers to improve classification accuracy while maintaining computational efficiency. A dataset of 3,114 annotated images representing three severity classes (mild damage, minimal or no damage, and severely damaged or dead) was used to train and evaluate the model. The custom model achieved the highest performance among seven architectures, with 95% accuracy, a compact model size of 1.37 MB, and an inference time of 15.4 seconds, outperforming state-of-the-art CNNs and Vision Transformer baselines. Per-class analysis confirmed the model’s robustness across all damage categories, with minor misclassifications primarily linked to poor lighting and visual symptom overlap. These results demonstrate the potential of the lightweight model for accurate, efficient, and scalable freeze damage detection, offering a practical and deployable solution for real-time decision-making in precision agriculture.