Evaluation of multispectral imaging for freeze damage assessment in strawberries using AI-based computer vision technology

• DOI: 10.1016/j.atech.2025.100851
• Publication Date: 2025-02-21T00:59:29Z
• Authors: Sunil GC; Amin Khan; David Horvath; Xin Sun
• Keywords: HD9000-9495; Agriculture (General); Machine learning; Computer vision; Deep learning; Agricultural industries; NDVI, Frost damage; S1-972

Climate change has increased the propensity for sudden temperature drops in early fall and late spring, resulting in greater chances for frost damage to agricultural crops and fruits. The development of frost-tolerant crops necessitates the application of artificial intelligence (AI) and computer vision technologies to make genetic studies of freezing damage consistent and less error prone. To advance the field, this study explored both feature engineering (feature extraction and feature selection) and deep learning (DL) approaches for frost damage classification. Feature engineering extracted various vegetative indices, while deep learning models, such as CNNs, automatically learned features from multispectral (RGNIR) images. Both machine learning (ML) and deep learning models were trained, validated, and tested on a dataset of RGNIR images acquired from strawberry plants cultivated in a greenhouse setting. For machine learning (ML) models, 80 % of the 493 images were used for training and 20 % for testing. For deep learning (DL) models, the data was divided into training, validation, and testing sets using a 70:15:15 ratio. Vegetative indices feature such as modified chlorophyll absorption ratio index (MCARI), modified transformed vegetation index (MTVI), and normalized difference vegetation index (NDVI) play significant role in frost damage classification in strawberry. Support vector machine (SVM) with backward feature elimination (F1-score: 87 %) outperformed other machine learning algorithms, including gradient boosting (GB), XGboost, and random forest (RF). Among deep learning models, DenseNet-169 achieved the highest F1-score of 93 %, surpassing DenseNet-121, DenseNet-201, Xception, Inception-ResNet-v2, Inception-v3, Spatial Attention, Channel Attention, and Spatial-Channel Attention. Studies shows that the integration of RGNIR images with computer vision and AI algorithms can be proven effective in classifying frost damage.