Quantifying freeze damage in strawberry plants using deep learning-based semantic segmentation

Traditional freezing damage quantification on plants often relies on human visual assessment, which can be inconsistent between individuals and even for the same person at different times. Additionally, this approach is time-consuming and labor-intensive. A computer system and application utilizing computer vision-based freezing-damage quantification artificial intelligence models have the potential to significantly streamline and improve the consistency of freezing damage assessment. This study focuses on the development of lightweight deep learning models with reduced numbers of parameters for freezing damage quantification. These models utilize semantic segmentation techniques to accurately classify and quantify background, dead tissue, and healthy tissue in plant images. Several deep learning image segmentation architectures, including DeepLabv3, UPerNet, PSPNet, U-Net, and SegNet, along with modification of UperNetResNet101 with attention mechanism and backbone architectures (densenet, efficientnet, convNeXt, and mobilenetv3) were trained on a greenhouse collected strawberry dataset. The UperNet model modified with a densenet backbone achieved the highest mean Intersection Over Union value of 0.77, surpassing the performance of the base model (0.76). Notably, the less complex ConvNeXt-tiny, model, which has 65.8 % fewer parameters than the UperNetResNet101 model, exhibited only a 2.63 % drop in performance, while improving fps by 53.88 %. This inference speed improvement can be attributed to a combination of factors, including variations in model complexity, model hardware compatibility, and efficient dataflow in model. Overall, testing of this AI-based model demonstrated its potential for freezing damage quantification in plant species beyond strawberries. • First-ever deep learning-based semantic segmentation of strawberry freeze damage. • Customized state-of-the-art models with backbone and attention enhancements. • Low-parameter model developed with minimal performance loss. • Upernet with densenet backbone achieved the best performance.