The dataset was collected from two separate experiments to increase the generalizability of the model. In experiment 1, the dataset was prepared by subjecting strawberry plants (cv. Honeoye) to various levels of freezing stress under controlled growth chamber conditions. Strawberry runners were planted in square 10 cm pots filled with horticulture potting media and initially grown in a greenhouse under controlled conditions at 20°C ± 2°C with a 12:12 light/dark photoperiod. When the plants had reached the 4‐leaf stage, the treatment groups were transferred to an acclimation chamber set at 5°C with a 12‐h photoperiod. A control set of plants was maintained at room temperature in the greenhouse. At the completion of each acclimation period (1, 2 and 3 weeks), the corresponding set of plants was transferred to a freezing chamber, where they were exposed to different freezing temperatures (−4°C, −8°C, or −12°C) for 4 h. Following freezing and a brief recovery period at 15°C, the plants were transferred back to the greenhouse and maintained under the controlled conditions (20°C ± 2°C with a 12:12 light/dark photoperiod). RGB images of plants were collected 3–7 days after the freezing treatments and return of plants to the greenhouse. In experiment 1, the plants were frozen at three temperature levels: slightly low (−4°C), moderately low (−8°C) and very low (−12°C) temperature. To address the gap in freezing temperature treatments of experiment 1 and obtain additional damage level images, another experiment was conducted using the same strawberry variety. In the second experiment, the protocols remained the same except for plants being exposed to a single freezing temperature of −10°C, a temperature determined to provide the greatest range of damage following freezing.

The RGB images of strawberry plants were collected using Canon EOS 90D digital cameras (Canon Inc., Tokyo, Japan) in natural lighting conditions in the greenhouse (Figure 1). In addition to these RGBs, images captured from the hyperspectral sensor (Specim FX10, Specim, Oulu, Finland) were added to increase the data size. Leveraging a large data size in DL enables the development of a robust model without model overfitting. There were a total of 759 image samples, which were labeled by experts into three categories: severely damaged or dead, mild damage, and minimal or no damage. The pie chart in Figure 2 illustrates the number of images in each category. Figure 3 demonstrates some examples of plant samples in each category.

Data preprocessing starts with cropping the images to minimize background content. Afterward, they were rescaled by dividing each pixel value by 255, normalizing the pixel values to a range between 0 and 1. Additionally, the images were resized to either 224 × 224 or 299 × 299 to ensure compatibility with certain DL architectures. These dimensions were chosen based on the standard input sizes for the selected convolutional neural network (CNN) architectures, ensuring optimal feature extraction and computational efficiency. The dataset was then split into training, validation, and test sets in a 70:15:15 ratio using a Python script. This resulted in 530 samples for training, 114 for validation, and 115 for testing (Table 1). To enhance the training data, augmentation techniques such as vertical and horizontal flipping, zooming, height and width shifting, shearing, and rotation were applied, effectively increasing the number of training samples. Data augmentation was performed using Keras ImageDataGenerator. Although the number of stored training images remained constant (530), random augmentation produced a new transformed version of each image at every epoch, resulting in approximately (epoch × training samples) augmented variants encountered by the model during training.

For model development, four of the most popular state‐of‐the‐art CNN architectures were selected: DenseNet‐121, Xception, ResNet‐50, and Inception v3. These models have been extensively utilized for plant stress detection tasks and have proven to perform effectively. For instance, Rezaei et al. (2024) employed all four models for disease detection in barley. Malvade et al. (2022) utilized Inception, ResNet‐50, and DenseNet for disease detection in paddy. Chandel et al. (2024) applied Inception and ResNet‐50 for water stress detection in wheat and maize crops. Sujatha et al. (2023) used DenseNet and ResNet for disease detection in paddy crops. These models have shown consistent performance across various plant stress detection tasks and agricultural applications. Their extensive use in previous research underscores their reliability and robustness, making them suitable choices for the freeze damage classification task in this study.

In this study, two approaches were used for model training: TL and training a model from scratch. Both methods were applied to assess whether a pretrained ImageNet model or a model trained on a custom strawberry dataset would provide better performance. The computational environment used for deep learning model training consisted of an AMD Ryzen 57,530 U processor with Radeon Graphics and 32 GB of RAM (Advanced Micro Devices Inc., Santa Clara, California, USA).

To classify freeze damage using TL, four models employed pretrained weights from the ImageNet dataset. Initially, the top fully connected layers of these models were removed, retaining only the convolutional base. The pretrained ImageNet weights were loaded, and the layers were frozen to maintain the features learned from ImageNet without updating them during training. The layers were then flattened to convert the 2D feature maps into a 1D feature vector, which is a necessary step before adding fully connected (dense) layers. A dense layer with three neurons (corresponding to the three classes) and a softmax activation function was subsequently added. Unlike TL models, scratch models started with randomly initialized weights, learning features and patterns directly from the training data. While TL offers advantages such as reduced computational cost and lower data requirements, it also has potential drawbacks. For example, if the pretrained model was trained on a dataset significantly different from the target task's dataset, it may not transfer effectively, leading to poor performance.

There were differences in choosing model parameters when training the model using these two approaches. In TL, the model has already been trained on a large dataset (such as ImageNet) and has learned meaningful features. In contrast, when training a model from scratch, the parameters are initialized randomly, requiring the model to learn all features and representations directly from the training data, which demands more data and longer training time. Table 2 presents the model training parameters for all models during both TL and scratch learning. The number of epochs is set to 500 for TL and 700 for scratch learning. These different epoch values were chosen to ensure optimal model performance during training. The higher number of epochs for scratch learning reflects the additional time required for training and optimization, which is expected since, in TL, all the backbone layers are pretrained and optimized.

Several evaluation metrics were employed to assess the classification models, including accuracy, precision, recall, and F1‐score (Sokolova and Lapalme 2009). Accuracy measures the proportion of correct predictions made by the model out of the total number of samples. Precision reflects the proportion of correctly predicted positive cases that were truly positive, calculated by dividing the number of true positives by the total number of predicted positives. Recall, on the other hand, indicates how well the model identified actual positive cases, determined by the ratio of correctly predicted positives to all actual positive cases. The F1‐score provides a balance between precision and recall, reaching its maximum value when precision equals recall.

Furthermore, the size of each model, their trainable parameters, and inference time were compared to assess model performance. Inference time refers to the amount of time it takes for a trained DL model to process new input data and make predictions. The inference time measurement for the models was processed on a 12th Gen Intel Core i7‐12650H (16 CPUs), ~2.7 GHz, and 16 GB of RAM; NVIDIA GeForce RTX 4050 Graphics Processor (NVIDIA Corporation, Santa Clara, California, USA). To ensure a reliable measurement, the prediction was run multiple times (500), and the average was considered the inference time of a model.

In this training process, the models (Densenet‐121, Inception V3, ResNet50, and Xception) were pretrained on the ImageNet dataset, and they contain the weights and biases learned from ImageNet. Table 3 shows that Inception V3 performs better than the other models, with an accuracy of 84%. Xception also provides good performance, with the best precision, recall, and F1‐score, and its accuracy is close to Inception. However, ResNet‐50 and DenseNet‐121 performed poorly compared to the other models, with an accuracy of 71%.

Figure 4 and Figure 5 demonstrate the confusion matrices for all TL models. DenseNet predicted 18 minimal or no damage samples as mild damage, which was less than the samples it predicted correctly (16) for this category (Figure 4a). Inception predicted seven minimal or no damage samples as mild damage (Figure 4b). ResNet‐50 model predicted 16 minimal or no damage samples as mild damage (Figure 5a). Xception predicted 14 minimal or no damage samples as mild damage (Figure 5b). In conclusion, the TL models mostly made wrong predictions for minimal or no damage category as they are wrongly predicted as mild damage.

Figure 6 and Figure 7 show the classification performance of TL model, such as precision, recall, and F1‐score for each category. Precision indicates how many of the positive predictions were actually correct. High precision shows that the model has a low false positive rate. For the severely damaged category, Xception has 100% precision, which means all of the positive predictions were actually positive. However, ResNet has a low precision rate (79%) for this category. For minimal or no damage category, ResNet has a 100% precision rate. The precision rate for mild damage category is comparatively lower than the other two classes, ranging from 43% to 56% for all TL models, indicating a large number of predicted mild damage cases are not actually mild damage plants. Recall shows how well the model can identify all positive instances. Figure 6 and Figure 7 show all models has good recall value (≥ 90%) for severely damaged category, where Inception captured all the positive instances (100%) correctly. However, for mild damage class Inception only shows 47% recall score. The F1‐score provides a balance between precision and recall. A high F1‐score indicates that the model performs well in terms of both precision and recall. It can be observed from Figure 6 and Figure 7 that Xception has the highest F1‐score (98%) for severely damaged category among all TL models. However, DenseNet has the highest (68%) F1‐score for mild damage class, and Inception has the highest F1‐score (83%) for minimal or no damage class.

Table 4 shows parameters of TL models, such as model size, number of total parameters, number of trainable parameters, and inference time. DenseNet‐121 is the smallest model of size 27.42 MB and with the lowest number of parameters, which is best for practical deployment on edge devices. ResNet‐50 has the largest size (91.13 MB), with Inception V3 and Xception being slightly smaller at 84.67 and 81.92 MB, respectively. All TL models have much fewer trainable parameters compared to total parameters, suggesting that much of its architecture is frozen for training efficiency. Xception achieved the fastest inference time, making it ideal for speed‐critical tasks.

In this training process, the models (Densenet‐121, Inception V3, ResNet50, and Xception) started with randomly initialized weights and learned features and patterns directly from the training data. ResNet‐50 outperformed other models with an accuracy of 97% on test data as shown in Table 5. However, DenseNet‐121 provided better recall than the other models, which is 96%.

Figure 8 and Figure 9 show the confusion matrix for four models that were trained from scratch. DenseNet predicted four minimal or no damage samples as mild damage and one sample as severely damaged as shown in Figure 8a. It predicted all 19 samples in mild damage correctly. It also performed very well in predicting severely damaged class as only one sample was wrongly predicted as mild damage. However, for Inception model, four minimal or no damage samples were predicted as severely damaged and two samples were predicted as mild damage (Figure 8b).

ResNet‐50 predicted one minimal or no damage class as mild damage, and one sample was predicted as severely damaged. However, it could predict all samples in the severely damaged class correctly (Figure 9a). Xception predicted five minimal or no damage samples as mild damage (Figure 9b).

Figure 10 and Figure 11 show the classification performance of scratch models, such as precision, recall, and F1‐score for each category. For severely damaged class, the models have good precision, ranging from 93% to 100%. For mild damage class, it ranges from 78% to 94%, whereas for minimal or no damage class, the precision ranges from 95% to 100% among the scratch models. Additionally, Inception and ResNet has 100% recall score for severely damaged class. DenseNet has 100% recall value for mild damage class. For F1‐score, severely damaged class performed better than the other categories, ranging from 96% to 99%. F1‐score has comparatively poor value for the mild damage class across all the scratch models.

Table 6 shows parameters of scratch models, such as model size, number of total parameters, number of trainable parameters, and inference time. DenseNet‐121 is the smallest model at 26.86 MB with the lowest number of parameters, which makes it highly efficient for deployment in environments with limited storage. ResNet‐50 has the largest size at 90 MB, followed closely by Inception V3 (83.19 MB). ResNet‐50 has the fastest inference time at 3.0016 s, which makes it efficient for real‐time or time‐sensitive tasks.

In the field of precision agriculture, several studies have compared the performance of models trained from scratch with those utilizing TL to evaluate and understand their relative strengths, weaknesses, and suitability for a specific task or dataset. Gulzar (2024) introduced a novel deep learning model for seed classification, comparing training the Xception model from scratch versus using TL with a pretrained Xception model. Xu et al. (2022) compared both TL and models trained from scratch while developing models for versatile plant disease recognition with limited data.

In this work, four deep learning (DL) models (Densenet‐121, Inception V3, ResNet50, and Xception) were evaluated using two training approaches: TL models with ImageNet weights and models trained from scratch. Evaluation of these two approaches indicates that the models trained from scratch on the strawberry dataset outperformed the TL models utilizing ImageNet weights. This demonstrates that the features learned from the ImageNet dataset are not sufficiently relevant for classifying freeze damage conditions in strawberry plants. Moreover, the training data is large and of high quality enough for the scratch models to outperform the TL models. A larger and more diverse dataset, encompassing a broader range of scenarios and variations, is crucial for training robust models. For example, a plant's appearance can change significantly within a few days following a freezing event. As the plant begins to recover, new tissue may develop, leading to variations in image appearance over time.

These results are associated with the findings of Gupta et al. (2024), where a transformer‐based model was trained in both ways, TL and from scratch. For TL, they first used ModelNet10 data to pretrain and then fine‐tuned the pretrained model to classify the MNIST dataset. For training from scratch, they trained the model from scratch on the MNIST dataset. They observed that transfer learned models did not outperform the scratch models since the two datasets have a large difference in the degree of the distributions. Similarly, He et al. (2018) stated that ImageNet pretraining does not necessarily help reduce overfitting unless the data size is significantly small. Additionally, Zhao et al. (2024) indicated that TL models may encounter a domain mismatch issue between the pretraining and target domains, leading to poor performance.

It is noteworthy that the ResNet‐50 model trained from scratch achieved the highest accuracy of 97%, outperforming the other models in this study. The superior performance of ResNet‐50 trained from scratch can be attributed to its residual learning structure, which facilitates efficient gradient propagation and better feature reuse across layers. Unlike Inception or Xception, which rely on multiscale or depthwise separable convolutions, ResNet's skip connections allow deeper feature learning without vanishing gradients. Consequently, it can effectively capture the complex texture variations associated with freeze damage symptoms in strawberry leaves. The findings of this study also align with prior research that compared various deep learning algorithms. For instance, Shin et al. (2021) evaluated AlexNet, SqueezeNet, GoogLeNet, ResNet‐50, SqueezeNet‐MOD1, and SqueezeNet‐MOD2 for detecting powdery mildew disease on strawberry plants and reported that ResNet‐50 achieved the highest classification accuracy of 98.11%. Similarly, Rajwade et al. (2024) compared ResNet‐50 and MobileNetV2 for water stress detection in Okra plants, with ResNet‐50 outperforming MobileNetV2. Malvade et al. (2022) compared Inception‐V3, VGG‐16, ResNet‐50, DenseNet‐121, and MobileNet‐28 and found that ResNet‐50 achieved the highest accuracy of 92.16%. Additionally, (Khan et al. 2024) evaluated VGGNet, InceptionV3, ResNet‐50, and InceptionResNetV2, with ResNet‐50 demonstrating superior accuracy of 87.51%. These consistent findings highlight ResNet‐50's effectiveness across diverse plant stress detection tasks.

A careful evaluation of the classification models for each category suggests that the models are predicting the severely damaged class more accurately than the other two categories. However, the models have difficulty in identifying the mild damage class as it sometimes mistakes a mildly damaged plant for a minimal or no damage plant. One of the reasons can be a lower number of mild damage samples than the other two categories. Furthermore, there is a subtle difference between the symptoms of some mild damage plants and a minimal damage plant to distinguish them properly. This issue is significant for the TL models as the features learned from the ImageNet dataset are not adequate for the models to identify these subtle differences. Scratch models have far better classification performance for the mild damage category than the TL models.

Model parameters, such as size and inference time, are critical metrics in evaluating the performance of a DL model, especially when it comes to practical deployment in real‐world applications. It was observed that DenseNet‐121 is the smallest model with the smallest number of parameters. This makes it ideal for memory‐constrained applications. However, it has the highest inference time, which makes it unsuitable for low‐latency tasks. This observation aligns with a previous statement by Jain (2024), who noted that during the inference of the DenseNet model, the concatenation operations can slow down the forward pass, increasing prediction latency. Moreover, Chao et al. (2019) stated that memory traffic generated by extracting the feature map of the middle part of DenseNet is a factor affecting the inference speed of the network model. Their results indicated DenseNet‐121 (51.5 ms) having higher inference time than ResNet‐50 (46.5 ms), even with a significantly lower number of parameters. In this study, ResNet‐50 trained from scratch had the lowest inference time (3.0016 s), making it the fastest model overall.

The findings of this study have significant implications for advancing both scientific research and practical applications in the field of precision agriculture. Employing computer vision and DL for freeze damage detection, this work highlights the potential of exploiting automated systems to address the challenges associated with traditional methods of plant stress assessment, which are often subjective, time‐consuming, and labor‐intensive. The adoption of DL‐based systems, as demonstrated in this study, can facilitate timely and accurate stress detection, enabling farmers to make informed decisions that minimize crop losses. Additionally, the comparison between TL and scratch learning demonstrates the value of developing custom datasets for agricultural applications. This reinforces the need for researchers to focus on generating high‐quality, domain‐specific data to enhance model performance and generalizability. The evaluation of multiple state‐of‐the‐art CNN architectures offers valuable insights into model selection, balancing tradeoffs between accuracy, inference speed, and hardware constraints. Furthermore, because the models are trained on RGB images, which can be easily captured using standard digital cameras or smartphones, the system offers a cost‐effective solution that can be readily adopted in both field and greenhouse environments. For practical deployment, cameras can be mounted on fixed stations, mobile robots, or drones to enable rapid, nondestructive assessment of freeze damage following cold events. The methods and findings presented in this study can serve as a reference for exploring similar applications across other crops and stress types, enabling the research community to replicate and extend this work.

The dataset used in this study showed an imbalance in class distribution, particularly with fewer samples in the mild damage category compared to the minimal or no damage and severely damaged categories. This imbalance may have contributed to the model's challenges in accurately classifying mild damage cases. Additionally, subtle differences between the phenotypic markers of mild damage and the minimal or no damage categories presented classification challenges for the models, especially those trained with TL. Future studies should concentrate on increasing the number of samples in underrepresented categories, such as mild damage. Techniques like Synthetic Minority Oversampling Technique (SMOTE) or Adaptive Synthetic (ADASYN) sampling can effectively balance the dataset. Collecting images in diverse environmental conditions, including outdoor fields and various growth stages, can enhance the generalizability and robustness of the models. Incorporating additional data modalities, such as thermal or multispectral imaging alongside RGB images, may help capture subtle differences in freeze damage symptoms and boost model performance. Future work should aim to optimize model architectures for faster inference without sacrificing accuracy, enabling real‐time applications in precision agriculture. The use of emerging deep learning architectures, such as Vision Transformers (ViT) or hybrid models that combine CNNs and attention mechanisms, could be explored to further enhance model performance.