|
Multi-Species Weed and Crop Classification Comparison Using Five Different Deep Learning Network Architectures
Sunil GC1, Yu Zhang1, Kirk Howatt2, Leon G. Schumacher1, Xin Sun1,*
Published in Journal of the ASABE 67(2): 275-287 (doi: 10.13031/ja.15590). Copyright 2024 American Society of Agricultural and Biological Engineers.
1Agricultural and Biosystems Engineering, North Dakota University, Fargo, North Dakota, USA.
2Department of Plant Sciences, North Dakota State University, Fargo, North Dakota, USA.
* Correspondence: xin.sun@ndsu.edu
Submitted for review on 13 March 2023 as manuscript number ITSC 15590; approved for publication as a Research Article and as part of the Artificial Intelligence Applied to Agricultural and Food Systems Collection by Associate Editor Dr. John Fulton and Community Editor Dr. Yiannis Ampatzidis of the Information Technology, Sensors, & Control Systems Community of ASABE on 13 December 2023.
Highlights
Abstract. The detection of individual weed and crop species from RGB images is a challenging task that becomes even more difficult as the number of species increases. This is because similarities in the phenotypic traits of weeds and crops make it difficult to accurately distinguish one species from another. In this study, five deep learning Convolutional Neural Networks (CNNs) were employed to classify six weed and eight crop species from North Dakota and assess the performance of each model for specific species from a single image. An automated data acquisition system was utilized to collect and process RGB images twice in a greenhouse setting. The first set of data was used to train the CNN models by updating all of its convolutional layers, while the second set was used to evaluate the performance of the models. The results showed that all CNN architectures, except Densenet, demonstrated strong performance, with macro average f1-scores (measurement of model accuracy) ranging from 0.85 to 0.87 and weighted average f1-scores ranging from 0.87 to 0.88. The presence of three weed classes—palmer amaranth, redroot pigweed, and waterhemp, all of which share similar phenotypic traits—negatively affected the model's performance. In conclusion, the results of this study indicate that CNN architectures hold great potential for classifying weed and crop species in North Dakota, with the exception of situations where plants have similar visible characteristics.
Weed detection is a crucial step in the development of machine vision-based chemical or mechanical weed control systems (Thompson et al., 1991). In vision-based mechanical weed control systems, multi-classification of weed and crop species is not essential, as the mechanical system applies the mechanical weeding tool regardless of weed species. However, in chemical weed control, the classification of weed and crop species can be beneficial for site-specific herbicide treatment. This is because the herbicide dose rate can be adjusted based on the weed species and biomass, thereby reducing herbicide usage (Thompson et al., 1991). The decrease in herbicide usage results in lower agriculture input costs and fewer risks to the environment and human health (Cech et al., 2022; Tudi et al., 2021). Therefore, an effective vision-based weed detection system can play a significant role in precision-based site-specific weed management.
The vision-based weed detection system uses computer vision and AI algorithms to identify weeds and crops within images. The recent progress in the field of deep learning in computer vision has shown promising results for classifying weed and crop species (Wu et al., 2021). Convolutional neural networks (CNNs), such as VGG (Simonyan and Zisserman, 2014), EfficientNet (Tan and Le, 2019), Densenet (Huang et al., 2016), and ConvNeXt (Liu et al., 2022), are frequently used in various computer vision tasks. These architectures can be used in agriculture to develop a weed classification system. However, training these models requires large datasets, which are limited for weed and crop detection (Alzubaidi et al., 2021; Kamilaris and Prenafeta-Boldú, 2018). To overcome these limitations, transfer learning and data augmentation techniques are commonly used (Alzubaidi et al., 2021; Hasan et al., 2021; Weiss et al., 2016). Transfer learning is a machine learning technique in which knowledge learned from a task is re-used in related tasks to boost a model’s performance, whereas data augmentation is a technique used to increase training data sets in machine learning using an existing dataset artificially.
 |
| Figure 1. Greenhouse image acquisition setup example for one bench with camera and plant (weed and crop) pots. One bench had Canon EOS 90D cameras and remaining bench had four Canon EOS T7 cameras; (a) Camera and (b) pot with promix, weed, and crop plant. |
Existing weed and crop classification models have been either for a single crop paired with weed species or only weed species alone, as seen in the review by Hasan et al. (2021). For example, Chen et al. (2022) developed a weeds-only model for cotton production system, Subeesh et al. (2022) developed a model for bell peppers without multi-weed classification, and Jiang et al. (2020) built a model for corn crop with four weeds. In cases where weed classification results are poor for certain weed species, crop classification can be used to identify weed areas in an image by excluding crop area using excess green image processing algorithm. The development of a comprehensive model for multiple crop production systems can save time and resources by eliminating the need to develop multiple models for different systems.
Furthermore, there are limited studies (Chen et al., 2022; Yu et al., 2019) that focus on the classification performance of multiple CNN models for a large number of crop or weed species. This has created a research gap in determining the best and worst-performing weed and crop classes when building a comprehensive weed classification model for multiple crop production systems. Additionally, understanding the best CNN architectures for specific types of weed and crop species is important. For example, if farmers are trying to control the waterhemp on their farm and have access to multiple models, they can use a model with the highest performance for waterhemp classification. Therefore, a comparison of individual crop and weed classification results for multiple CNN models is crucial for the development of a targeted weed control system.
The objective of this study was (1) to train all layers of five CNN architectures using weed and image datasets, and compare models for six weed species and eight crop species based in North Dakota, U.S.; (2) to determine the best and worst-performing weed and crop classes for the five CNN models; and (3) to build a single comprehensive CNN multi-classification model.
Materials and Methods
Weed and Crop Experiment Design and Image Acquisition
The study was performed at North Dakota State University's Waldron greenhouse, where weeds and crops were grown in pots filled with Promix mixture in the spring of 2021. Four benches, each with an approximate area of 48.8 ft2 (figure 1), were used to plant six weed species (horseweed: Conyza canadensis, kochia: Bassia scoparia, palmer amaranth: Amaranthus palmeri, ragweed: Ambrosia artemisiifolia, redroot pigweed: Amaranthus retroflexus, and waterhemp: Amaranthus tuberculatus) and eight crop species (blackbean: Phaseolus vulgaris, canola: Brassica napus, corn: Zea mays, field pea: Pisum sativum, flax: Linum usitatissimum, lentil: Lens culinaris, soybean: Glycine max, and sugar beet: Beta vulgaris) (figure 2). These species were selected because all eight crops are grown in North Dakota and five of the weed species (excluding palmer amaranth) are prevalent in the state (Bryson and DeFelice, 2010). However, palmer amaranth is moving north from southern states and has already been introduced in North Dakota (Oliveira et al., 2022). Each of the four benches was planted with two crop species and six weed species, resulting in 72 pots for each weed species and 63 pots for each crop species.
  | ||||
| (a) Crop species | (b) Weed species | |||
| Figure 2. The RGB image of weed and crop species used for the study. These images were from set 1 greenhouse data collection. | ||||
After weed and crop planting, a data acquisition system was set up in the greenhouse to collect weed and crop RGB images. In this study, twelve Canon EOS T7 and three EOS 90D visible spectrum cameras (Canon Inc., Tokyo, Japan) were deployed in the greenhouse to capture images from each bench. All cameras were mounted in the fixed position on all four benches, with an approximate height of 48±2 inches from camera to bench with pot. The camera captures images at intervals of 30 minutes all the time unless greenhouse power goes off during the night. Planting was performed twice to collect weed and crop images, which were used for building artificial intelligence models and model evaluation. The first round of data collection was performed from 18 March to 2 April in 2021, which was termed “Set 1,” whereas the second round of data collection was performed from 16 April to 3 May in 2021, which was termed “Set 2.” Set 1 data was used for building artificial intelligence models, and set 2 data was used to evaluate the models from set 1 data.
Data Cleaning, Data Labeling, and Data Description
The automatic data acquisition system built using a digital camera, Linux desktop, and cloud was used to capture weed and crop images from all four benches. Data cleaning was performed to remove blurred images and images with noise (human body parts, water pipe, and water during watering) above weed or crop plants. After cleaning, labeling of an image was performed using LabelImg (https://github.com/?heartexlabs/labelImg) software. Images were labeled, which saves labeled weed and crop objects in an Extensible Markup Language (XML) file format with object bounding box information. Image labeling was performed on a single image of each day’s image collection from each camera, similar to our previous study (GC et al., 2022).
| Table 1. The total number of training, validation, and testing images of crop and weed species for set 1 data collection and total images from set 2 data collection as testing images. | ||||
| Weed and Crop | Set 1- training | Set 1- validation | Set 1- testing | Set 2- testing |
| blackbean | 16257 | 3484 | 3484 | 15119 |
| canola | 8936 | 1916 | 1916 | 15131 |
| corn | 15803 | 3387 | 3387 | 17573 |
| field pea | 12816 | 2747 | 2747 | 19617 |
| flax | 14353 | 3076 | 3076 | 14526 |
| horseweed | 10961 | 2349 | 2349 | 19394 |
| kochia | 8997 | 1929 | 1929 | 13218 |
| lentil | 17514 | 3753 | 3753 | 12191 |
| palmer amaranth | 11790 | 2527 | 2527 | 10990 |
| ragweed | 16113 | 3454 | 3454 | 19559 |
| redroot pigweed | 7701 | 1651 | 1651 | 6077 |
| soybean | 11068 | 2372 | 2372 | 13773 |
| sugar beet | 9168 | 1965 | 1965 | 14301 |
| waterhemp | 20611 | 4418 | 4418 | 18068 |
| Total | 182088 | 39028 | 39028 | 209537 |
The next step after the labeling was image cropping using the labeled image bounding box coordinates for weed and crop classes. For single-day image cropping, each day’s single labeling file was used for the remaining images of that day from the same camera. This approach is similar to the approach applied to our previous study (GC et al., 2022). This was possible due to the fixed position of cameras, which reduced the human manual work involved in image cropping. Image cropping from all camera images was performed using a Python script for both set 1 and set 2 datasets.
Crops and Weed Image Classification Using a Convolutional Neural Network
There were a total of 14 classes of weed and crop species for the deep-learning image classification task. There were 260,144 set 1 and 209,537 set 2 images after cropping into a single image of each weed and crop species. Table 1 shows the number of each class for weed and crop species. The deep learning classification models were built from the set 1 dataset, and the classification models thus obtained were tested on unfamiliar images of the set 2 labeled data. During the model training, set 1 data were divided into the training, validation, and testing data sets in the ratio of 70:15:15. The total number of training, validation, and testing dataset images from set 1 datasets were 182,088, 39,028, and 39,028, respectively (table 1), whereas the total number of set 2 testing images was 209,537. The number of images in each class is not balanced because some plants did not grow well, and the number of crop and weed species pots was 63 and 72 for each class, respectively. The training datasets were augmented to increase training data size using shift, flip, rotation, and zoom, which is believed to reduce model overfitting and increase model robustness (Alzubaidi et al., 2021).
Five deep learning convolutional neural network (CNN) architectures, Xception (Chollet, 2016), Densenet (Huang et al., 2016), MobileNetV3 (Howard et al., 2019), EfficientNet (Tan and Le, 2019), and ConvNeXt (Liu et al., 2022), were used to build five deep learning weed and crop classification models. These models were chosen specifically because of the variation of convolutional and pooling layers, which determines the accuracy and inference time of the architecture (Alzubaidi et al., 2021). Besides this, ConvNeXt is constructed from standard convolutional network (ConvNet) modules, which outperformed the transformers (natural language processing dominant backbone architecture) in terms of accuracy, robustness, and scalability across the deep learning benchmark datasets (Liu et al., 2022). In this study, the transfer learning approach was not used due to the sufficient size of data to train all the trainable layers parameters of CNN architecture (Espejo-Garcia et al., 2020b; GC et al., 2022). The output layer was modified with the number of weed and image classes (14) used in this study. Categorical cross entropy and adam optimizer (Kingma and Ba, 2015) were used as the loss function and optimizer, respectively (table 2). Out of the five CNN architectures used, ConvNeXt had the highest number of parameters, whereas Densenet had the lowest number of parameters, which is clearly depicted in table 2. Furthermore, MobileNetV3Large was a variant of the MobileNetV3 model with increased depth and higher resource requirements, in contrast to MobileNetV3Small (Howard et al., 2019). Similarly, EfficientNet had several variations, among which EfficientNet-b0 served as the fundamental model for all the other variants from b1-b7 (Tan and Le, 2019). The total number of epoch and batch size for training all five architectures were 500 and 32, respectively, with 178 iterations on each epoch.
| Table 2. The total number of model parameters, batch size, loss function, and optimizer used. | ||||
| Model | Number of Parameters | Batch Size | Loss Function | Optimizer |
| Xception | 20.92M | 32 | categorical cross entropy | adam |
| Densenet | 0.20M | 32 | categorical cross entropy | adam |
| MobileNetV3Large | 4.24M | 32 | categorical cross entropy | adam |
| EfficientNetb0 | 4.07M | 32 | categorical cross entropy | adam |
| ConvNeXt | 27.8M | 32 | categorical cross entropy | adam |
The model performance was evaluated using four metrics: (1) accuracy, (2) precision, (3) recall, and (4) f1-score. In addition, a normalized confusion matrix over the actual label was obtained. For the deep learning model training tensorflow (tensorflow.org) and keras (keras.io), the Python application program interface (API) was used, whereas for performance metric evaluation, the Scikit-learn Python API was used (Pedregosa et al., 2011). All five architectures were trained with a GeForce GTX 1080 Ti 11GB Nvidia (Nvidia Co., Santa Clara, USA) graphical processing unit (GPU).
| Table 3. Xception weed and crop classification model performance evaluation in terms of precision, recall, and f1-score for eight crop species and six weed species. | |||||||||||
| Precision | Recall | F1-score | Number of Test Data | ||||||||
| Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | ||||
| blackbean | 0.99 | 0.96 | 1.00 | 0.97 | 1.00 | 0.97 | 3484 | 15119 | |||
| canola | 1.00 | 0.97 | 1.00 | 0.97 | 1.00 | 0.97 | 1916 | 15131 | |||
| corn | 1.00 | 0.95 | 1.00 | 0.99 | 1.00 | 0.97 | 3387 | 17573 | |||
| field pea | 1.00 | 0.91 | 1.00 | 0.99 | 1.00 | 0.95 | 2747 | 19617 | |||
| flax | 1.00 | 0.93 | 1.00 | 0.86 | 1.00 | 0.89 | 3076 | 14526 | |||
| horseweed | 1.00 | 0.99 | 1.00 | 0.98 | 1.00 | 0.98 | 2349 | 19394 | |||
| kochia | 0.99 | 0.91 | 1.00 | 0.96 | 1.00 | 0.94 | 1929 | 13218 | |||
| lentil | 1.00 | 0.99 | 1.00 | 0.90 | 1.00 | 0.95 | 3753 | 12191 | |||
| palmer amaranth | 0.96 | 0.46 | 0.99 | 0.69 | 0.98 | 0.55 | 2527 | 10990 | |||
| ragweed | 1.00 | 1.00 | 1.00 | 0.98 | 1.00 | 0.99 | 3454 | 19559 | |||
| redroot pigweed | 0.97 | 0.39 | 0.99 | 0.64 | 0.98 | 0.48 | 1651 | 6077 | |||
| soybean | 1.00 | 0.94 | 1.00 | 0.97 | 1.00 | 0.95 | 2372 | 13773 | |||
| sugar beet | 1.00 | 0.91 | 1.00 | 0.95 | 1.00 | 0.93 | 1965 | 14301 | |||
| waterhemp | 1.00 | 0.84 | 0.97 | 0.32 | 0.98 | 0.47 | 4418 | 18068 | |||
| macro avg | 0.99 | 0.87 | 0.99 | 0.87 | 0.99 | 0.86 | 39028 | 209537 | |||
| weighted avg | 0.99 | 0.90 | 0.99 | 0.88 | 0.99 | 0.88 | 39028 | 209537 | |||
(1)
(2)
(3)
(4)
where
TP= True positive
TN= True negative
FP= False positive
FN= False negative.
Results
Weed and Crop Classification Model Performance Evaluation
Xception Weed and Crop Classification Model Evaluation
The Xception convolutional neural network (CNN) model was trained on set 1 dataset, which was tested on both set 1 test datasets and set 2 test datasets. The set 2 test dataset was a complete data set from a different experiment in the greenhouse than that of the set 1 experiment. The classification results of all 14 classes in terms of precision, recall, and f1-score are depicted in table 3. The macro and weighted f1-score averages for set 1 test data were both 0.99. However, a macro and weighted f1-score for set 2 test data declined to 0.86 and 0.88, respectively, as expected due to being set 2 data from a completely different experiment than set 1 data. For set 1 test data, the lowest model performance was 0.98 for palmer amaranth, redroot pigweed, and waterhemp. For set 2 test data, f1-score values for palmer amaranth, redroot pigweed, and waterhemp were 0.55, 0.48, and 0.47, respectively, which was due to lower precision and recall values (table 3). However, for the remaining weed and crop classes, the f1-score was between 0.89 and 0.98. This decline in model performance on set 2 test data was mainly attributed to three weed species (palmer amaranth, redroot pigweed, and waterhemp) due to having similar phenotypic traits in these species (Bensch et al., 2003).
 |
| Figure 3. Normalized confusion matrix for Xception model evaluation on set 2 data. (BB: blackbean, Can: canola, FP: field pea, HW: horseweed, PA: palmer amaranth, RW: ragweed, RRPW: redroot pigweed, Soy: soybean, Sub: sugar beet, WH: waterhemp). |
The decline in the Xception model performance on set 2 test data was further elucidated by the confusion matrix in figure 3. The performance accuracy of palmer amaranth, redroot pigweed, and waterhemp was 69%, 63.9%, and 32.1%, respectively. When visualizing the confusion matrix, 20.3% of palmer amaranth images were detected as redroot pigweed, which was the highest percentage of misclassification for palmer amaranth. The second highest misclassification was contributed by waterhemp, with 4.78% of misclassification. When evaluating the confusion matrix for redroot pigweed, the highest and second highest number of redroot pigweed images were misclassified as palmer amaranth and soybean, with the misclassification percentage of 25.4% and 3.47%, respectively. Finally, when visualizing the waterhemp confusion matrix, 36.6% and 19.6% of the waterhemp images were misclassified as palmer amaranth and redroot pigweed, respectively. These results also show a higher shared misclassification between palmer amaranth and redroot pigweed. For waterhemp, most of the images were misclassified as palmer amaranth and redroot pigweed, but the reverse was not true. This resulted in a higher precision value and less recall value of waterhemp in Set 2.
Densenet Weed and Crop Classification Model Evaluation
The Densenet CNN model trained on set 1 training and validation datasets was evaluated with set 1 and set 2 test datasets. Table 3 depicts the precision, recall, f1-score, and number of test images for the Densenet model. The overall macro average and weighted average f1-score for the set 1 test dataset were 0.94 and 0.95, respectively, whereas the overall macro average and weighted average f1-score for the set 2 dataset were 0.76 and 0.78, respectively. The range of f1-score values for the set 1 test data was between 0.77 (redroot pigweed) and 1.00 (ragweed). There was a decline in f1-score values for set 2 test data, which was mostly attributed to waterhemp, redroot pigweed, and palmer amaranth classes with f1-score values of 0.25, 0.32, and 0.44, respectively. The weightage of the precision and recall for these three weed species to lower their respective f1-scores is depicted in table 4. The f1-score for the remaining weed and crop classes for the set 2 test data was between 0.79 (sugar beet) and 0.95 (ragweed).
The sharp decline of the model performance on set 2 test data was further elucidated by the confusion matrix in figure 4. The top three misclassified classes were waterhemp, palmer amaranth, and redroot pigweed, with classification accuracy of 15.5%, 38.5%, and 45%, respectively. However, classification accuracy for the remaining classes was much higher, between 79.7% (blackbean) and 99% (corn). When visualizing the confusion matrix for waterhemp, 24.4%, 16.8%, 14.5%, and 12.5% of waterhemp images were predicted as redroot pigweed, sugar beet, palmer amaranth, and corn, respectively. In addition, for palmer amaranth, 29.8%, 13.2%, and 6.61% of images were predicted as redroot pigweed, soybean, and waterhemp, respectively. Finally, for redroot pigweed, 19.7%, 18.3%, and 4.71% were predicted as palmer amaranth, canola, and soybean, respectively. This shows higher misclassification rates between palmer amaranth and redroot pigweed compared to crops and weeds with dissimilar phenotypic traits. Moreover, the misclassification rates were higher for waterhemp when confused with redroot pigweed and sugar beet, compared to those between crops and weeds with dissimilar phenotypic traits.
| Table 4. Densenet weed and crop classification model performance evaluation in terms of precision, recall, and f1-score for eight crop species and six weed species. | |||||||||||
| Precision | Recall | F1-score | Number of Test Data | ||||||||
| Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | ||||
| blackbean | 0.97 | 0.87 | 0.97 | 0.80 | 0.97 | 0.83 | 3484 | 15119 | |||
| canola | 0.86 | 0.83 | 1.00 | 0.98 | 0.93 | 0.90 | 1916 | 15131 | |||
| corn | 0.88 | 0.68 | 1.00 | 0.99 | 0.93 | 0.81 | 3387 | 17573 | |||
| field pea | 1.00 | 0.91 | 0.99 | 0.97 | 0.99 | 0.94 | 2747 | 19617 | |||
| flax | 0.99 | 0.92 | 0.98 | 0.81 | 0.99 | 0.86 | 3076 | 14526 | |||
| horseweed | 0.99 | 0.99 | 0.97 | 0.87 | 0.98 | 0.93 | 2349 | 19394 | |||
| kochia | 0.96 | 0.79 | 0.98 | 0.93 | 0.97 | 0.85 | 1929 | 13218 | |||
| lentil | 0.99 | 0.93 | 0.98 | 0.81 | 0.99 | 0.87 | 3753 | 12191 | |||
| palmer amaranth | 0.83 | 0.52 | 0.84 | 0.39 | 0.83 | 0.44 | 2527 | 10990 | |||
| ragweed | 0.99 | 0.93 | 1.00 | 0.97 | 1.00 | 0.95 | 3454 | 19559 | |||
| redroot pigweed | 0.87 | 0.25 | 0.69 | 0.45 | 0.77 | 0.32 | 1651 | 6077 | |||
| soybean | 0.98 | 0.91 | 0.96 | 0.82 | 0.97 | 0.86 | 2372 | 13773 | |||
| sugar beet | 0.94 | 0.69 | 0.97 | 0.92 | 0.95 | 0.79 | 1965 | 14301 | |||
| waterhemp | 0.97 | 0.68 | 0.89 | 0.16 | 0.93 | 0.25 | 4418 | 18068 | |||
| macro avg | 0.95 | 0.78 | 0.94 | 0.78 | 0.94 | 0.76 | 39028 | 209537 | |||
| weighted avg | 0.95 | 0.81 | 0.95 | 0.80 | 0.95 | 0.78 | 39028 | 209537 | |||
 |
| Figure 4. Normalized confusion matrix for Densenet model evaluation on set 2 data. (BB: blackbean, Can: canola, FP: field pea, HW: horseweed, PA: palmer amaranth, RW: ragweed, RRPW: redroot pigweed, Soy: soybean, Sub: sugar beet, WH: waterhemp). |
MobileNetV3 Weed and Crop Classification Model Evaluation
The MobileNetV3 model trained on set 1 data was tested with set 1 test data and set 2 data, whose performance evaluation is depicted in table 5 in terms of precision, recall, and f1-score. The overall macro and weighted f1-score averages for the set 1 test data were both 0.98, which was higher than the set 2 data with macro and weighted average f1-score values of 0.85 and 0.87, respectively. The f1-score values of the 14 classes for set 1 test data were between 0.91 (redroot pigweed) and 1.00 (canola, corn, field pea, and ragweed). The contribution of the individual class for the set 2 data overall f1-score is shown in table 5, where the highest contribution was from blackbean, canola, and ragweed with an f1-score value of 0.98. The three lowest f1-score classes for set 2 data were redroot pigweed, waterhemp, and palmer amaranth, with f1-score values of 0.42, 0.48, and 0.52, respectively. These lower f1-score values were attributed to low precision and recall values, which are depicted in table 5.
| Table 5. MobileNetV3 weed and crop classification model performance evaluation in terms of precision, recall, and f1-score for eight crop species and six weed species. | |||||||||||
| Precision | Recall | F1-score | Number of Test Data | ||||||||
| Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | ||||
| blackbean | 1.00 | 0.97 | 0.99 | 0.99 | 0.99 | 0.98 | 3484 | 15119 | |||
| canola | 1.00 | 0.98 | 1.00 | 0.99 | 1.00 | 0.98 | 1916 | 15131 | |||
| corn | 0.99 | 0.98 | 1.00 | 0.97 | 1.00 | 0.97 | 3387 | 17573 | |||
| field pea | 1.00 | 0.93 | 0.99 | 0.97 | 1.00 | 0.95 | 2747 | 19617 | |||
| flax | 1.00 | 0.98 | 0.99 | 0.84 | 0.99 | 0.90 | 3076 | 14526 | |||
| horseweed | 0.99 | 0.95 | 1.00 | 0.96 | 0.99 | 0.95 | 2349 | 19394 | |||
| kochia | 0.99 | 0.89 | 0.99 | 0.96 | 0.99 | 0.92 | 1929 | 13218 | |||
| lentil | 1.00 | 0.99 | 0.97 | 0.83 | 0.98 | 0.90 | 3753 | 12191 | |||
| palmer amaranth | 0.88 | 0.44 | 0.97 | 0.63 | 0.92 | 0.52 | 2527 | 10990 | |||
| ragweed | 1.00 | 0.97 | 1.00 | 0.99 | 1.00 | 0.98 | 3454 | 19559 | |||
| redroot pigweed | 0.91 | 0.33 | 0.91 | 0.58 | 0.91 | 0.42 | 1651 | 6077 | |||
| soybean | 1.00 | 0.97 | 0.98 | 0.96 | 0.99 | 0.96 | 2372 | 13773 | |||
| sugar beet | 0.97 | 0.86 | 1.00 | 0.96 | 0.99 | 0.90 | 1965 | 14301 | |||
| waterhemp | 0.99 | 0.74 | 0.97 | 0.35 | 0.98 | 0.48 | 4418 | 18068 | |||
| macro avg | 0.98 | 0.86 | 0.98 | 0.86 | 0.98 | 0.85 | 39028 | 209537 | |||
| weighted avg | 0.98 | 0.89 | 0.98 | 0.87 | 0.98 | 0.87 | 39028 | 209537 | |||
The misclassification of the 14 classes in the set 2 data is clearly visible in the figure 5 confusion matrix. The classification accuracy was between 35.4% (waterhemp) and 99.3% (blackbean). When visualizing the confusion matrix, the classification accuracy for waterhemp, redroot pigweed, and palmer amaranth were 35.4%, 58.1%, and 62.7%, respectively. Waterhemp had the lowest accuracy because 34.8% of waterhemp images were classified as palmer amaranth, while 20.1% of waterhemp images were misclassified as redroot pigweed. However, the f1-score calculation uses precision as well, which was increased due to less misclassification of other classes as waterhemp. The second lowest performing class was redroot pigweed, which was mainly due to the prediction of 28.8% of the redroot pigweed images as palmer amaranth. Finally, the third lowest performing class was palmer amaranth, which was mainly due to the misclassification of 25.9% of palmer amaranth images as redroot pigweed. This also shows the higher misclassification in the MobileNetV3 model for similarly visible phenotypic trait classes.
 |
| Figure 5. Normalized confusion matrix for MobileNetV3 model evaluation on set 2 data. (BB: blackbean, Can: canola, FP: field pea, HW: horseweed, PA: palmer amaranth, RW: ragweed, RRPW: redroot pigweed, Soy: soybean, Sub: sugar beet, WH: waterhemp). |
EfficientNet Weed and Crop Classification Model Evaluation
The EfficientNet model performance evaluation in terms of precision, recall, and f1-score for set 1 and set 2 data is depicted in table 6. The macro and weighted f1-score averages for set 1 data were both 0.99, which declined to 0.87 and 0.88, respectively, for set 2 data. When analyzing the f1-score for the set 1 data, all classes have an f1-score of 1.00 except redroot pigweed, palmer amaranth, and waterhemp, with f1-score values of 0.97, 0.98, and 0.99, respectively. However, for the set 2 data, the f1-score values for all 14 classes declined to a low of 0.52 (waterhemp) and a high of 0.98 (blackbean, canola, and ragweed). The three worst-performing classes for the set 2 data were waterhemp, palmer amaranth, and redroot pigweed, with f1-score values of 0.52, 0.54, and 0.57, respectively, which was attributed to lower precision and recall values as shown in table 6.
The set 2 data low performing classes and high performing classes classification accuracy and misclassification is depicted in figure 6 confusion matrix. The classification accuracy was between 99.2% (corn) and 32.1% (waterhemp). The lowest f1-score in waterhemp was mainly due to the misclassification of 36.6% and 19.6% of waterhemp as palmer amaranth and redroot pigweed, respectively. The next lowest performing class was redroot pigweed, which was mainly due to the misclassification of 25.4% of redroot pigweed as palmer amaranth. Finally, the third lowest performing class was palmer amaranth, which was mainly due to the misclassification of palmer amaranth as redroot pigweed. This shows similarity in how phenotypic traits between redroot pigweed, palmer amaranth, and waterhemp contributed to the low performance for set 2 data testing.
ConvNeXt Weed and Crop Classification Model Evaluation
The ConvNeXt classification results in terms of precision, recall, and f1-score for set 1 and set 2 data are illustrated in table 7, where set 2 data performance results were lower than set 1 data. The macro and the weighted f1-score average for set 1 data were 0.99 and 1.00, which were declined for set 2 data with values of 0.85 and 0.87, respectively. When analyzing the f1-score for set 1 data, all the classes except palmer amaranth, redroot pigweed, and waterhemp had the highest f1-score value of 1.00. The f1-score values were declined for all 14 classes when testing with set 2 data, which were between 0.49 (palmer amaranth) and 0.99 (ragweed). The top three low-performing classes for set 2 data were palmer amaranth, waterhemp, and redroot pigweed, with f1-score of 0.49, 0.50, and 0.53, respectively, which were also the top three set 2 data low performing classes for the previously discussed four deep learning CNN models. The precision and recall values shown in table 7 justify the reason behind the low f1-score in low-performing classes.
In order to go beyond precision and recall finding the cause of low-performing classes for set 2 data, the confusion matrix in figure 7 depicts a clear picture. In the confusion matrix for set 2 data, waterhemp, palmer amaranth, and redroot pigweed were the classes with low classification accuracy. In waterhemp, low classification accuracy was mainly due to the misclassification of 22% of palmer amaranth as waterhemp and 16.6% of waterhemp as redroot pigweed. Besides this, the prediction of waterhemp as kochia, sugar beet, and soybean also contributed, with a percentage between 4% and 8%. The next class contributing to low performance was palmer amaranth, in which 20.3% and 16.6% of the images were misclassified as redroot pigweed and waterhemp, respectively. The final one of the main contributing classes for the ConvNeXt model was redroot pigweed, whose 10.9% and 8.08% of images were misclassified as palmer amaranth, and waterhemp, respectively. This shows model performance decline in set 2 data was mainly contributed by palmer amaranth, waterhemp, and redroot pigweed, which was also the same for the Xception, Densenet, MobileNetV3, and EfficientNet models.
| Table 6. EfficientNet weed and crop classification model performance evaluation in terms of precision, recall, and f1-score for eight crop species and six weed species. | |||||||||||
| Precision | Recall | F1-score | Number of Test Data | ||||||||
| Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | ||||
| blackbean | 1.00 | 0.97 | 1.00 | 0.99 | 1.00 | 0.98 | 3484 | 15119 | |||
| canola | 1.00 | 0.97 | 1.00 | 0.99 | 1.00 | 0.98 | 1916 | 15131 | |||
| corn | 1.00 | 0.99 | 1.00 | 0.96 | 1.00 | 0.97 | 3387 | 17573 | |||
| field pea | 1.00 | 0.98 | 0.99 | 0.94 | 1.00 | 0.96 | 2747 | 19617 | |||
| flax | 0.99 | 0.90 | 1.00 | 0.96 | 1.00 | 0.93 | 3076 | 14526 | |||
| horseweed | 1.00 | 0.99 | 1.00 | 0.95 | 1.00 | 0.97 | 2349 | 19394 | |||
| kochia | 1.00 | 0.88 | 1.00 | 0.95 | 1.00 | 0.92 | 1929 | 13218 | |||
| lentil | 1.00 | 0.95 | 1.00 | 0.96 | 1.00 | 0.95 | 3753 | 12191 | |||
| palmer amaranth | 0.97 | 0.52 | 0.98 | 0.57 | 0.98 | 0.54 | 2527 | 10990 | |||
| ragweed | 1.00 | 0.97 | 1.00 | 0.99 | 1.00 | 0.98 | 3454 | 19559 | |||
| redroot pigweed | 0.99 | 0.49 | 0.95 | 0.69 | 0.97 | 0.57 | 1651 | 6077 | |||
| soybean | 1.00 | 0.92 | 1.00 | 0.97 | 1.00 | 0.94 | 2372 | 13773 | |||
| sugar beet | 0.99 | 0.84 | 1.00 | 0.97 | 1.00 | 0.90 | 1965 | 14301 | |||
| waterhemp | 0.99 | 0.70 | 0.99 | 0.41 | 0.99 | 0.52 | 4418 | 18068 | |||
| macro avg | 0.99 | 0.86 | 0.99 | 0.88 | 0.99 | 0.87 | 39028 | 209537 | |||
| weighted avg | 0.99 | 0.89 | 0.99 | 0.89 | 0.99 | 0.88 | 39028 | 209537 | |||
Five Deep Learning Weed and Crop Image Classification Comparisons
Five deep learning (DL) convolution neural network (CNN) weed, and crop classification models tested on 209537 images of set 2 data were compared with a box plot plotted based on individual class f1-score in figure 8. In the box plot, the median line of the Densenet box plot lies outside of the box of the remaining four models box plot below the f1-score value of 0.90, which shows the difference in the model performance in terms of f1-score. In addition, the median for the remaining four models is near each other, in between the f1-score value of 0.90 and 0.95. Besides this, a longer whisker of the Densenet box plot shows a high dispersion of f1-score values for an individual class. However, MobileNetV3 and EfficientNet have shorter whiskers and less dispersed outliers, showing less dispersion of the f1-score values of individual classes. This shows a high scattered f1-score value in Densenet and a low scattered f1-score value in EfficientNet. All five models had three outliers, which was due to the low performance in three weed classes (palmer amaranth, redroot pigweed, waterhemp), as discussed in the previous 3.1 section. The three outliers have caused the mean value of the model to be lower than the median value. The mean f1-score values for the model range from 0.76 for DenseNet to 0.87 for EfficientNet, which implicitly shows the best and worst-performing models, respectively.
Weed detection of individual weed species is also needed to develop precision weed control systems because of the different natures of weeds. In this study, a comparison of five CNN models on each weed and crop species classification was performed, which is depicted in figure 9 for weed species and figure 10 for crop species classification. For weed species, horseweed, kochia, and ragweed performed well, but palmer amaranth, redroot, and waterhemp performed poorly, which was due to the similar phenotypic traits in these underperformed weed species. In all six weed species, Densenet was the lowest-performing model, however, the best-performing models were different. The Xception model was the best-performing model for horseweed, kochia, and palmer amaranth, and the EfficientNet model was the best-performing model for redroot pigweed and waterhemp. In addition, the ConvNeXt model performed better than Xception, Densenet, and MobileNetV3 in redroot pigweed and waterhemp. In all five CNN models, ragweed was the best-performing class, whereas redroot pigweed was the worst-performing class in the MobileNetV3 model, palmer amaranth was the worst-performing class in the ConvNeXt model, and waterhemp was the worst-performing class in the remaining three CNN models.
 |
| Figure 6. Normalized confusion matrix for EfficientNet model evaluation on set 2 data. (BB: blackbean, Can: canola, FP: field pea, HW: horseweed, PA: palmer amaranth, RW: ragweed, RRPW: redroot pigweed, Soy: soybean, Sub: sugar beet, WH: waterhemp). |
| Table 7. ConvNeXt weed and crop classification model performance evaluation in terms of precision, recall, and f1-score for eight crop species and six weed species. | |||||||||||
| Precision | Recall | F1-score | Number of Test Data | ||||||||
| Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | Set 1 | Set 2 | ||||
| blackbean | 1.00 | 0.97 | 1.00 | 0.99 | 1.00 | 0.98 | 3484 | 15119 | |||
| canola | 1.00 | 0.98 | 1.00 | 0.97 | 1.00 | 0.98 | 1916 | 15131 | |||
| corn | 1.00 | 0.98 | 1.00 | 0.95 | 1.00 | 0.96 | 3387 | 17573 | |||
| field pea | 1.00 | 0.93 | 1.00 | 0.99 | 1.00 | 0.96 | 2747 | 19617 | |||
| flax | 1.00 | 0.97 | 1.00 | 0.88 | 1.00 | 0.92 | 3076 | 14526 | |||
| horseweed | 1.00 | 0.99 | 1.00 | 0.95 | 1.00 | 0.97 | 2349 | 19394 | |||
| kochia | 1.00 | 0.84 | 1.00 | 0.96 | 1.00 | 0.90 | 1929 | 13218 | |||
| lentil | 1.00 | 0.93 | 1.00 | 0.95 | 1.00 | 0.94 | 3753 | 12191 | |||
| palmer amaranth | 0.99 | 0.52 | 0.97 | 0.47 | 0.98 | 0.49 | 2527 | 10990 | |||
| ragweed | 1.00 | 0.99 | 1.00 | 0.99 | 1.00 | 0.99 | 3454 | 19559 | |||
| redroot pigweed | 0.97 | 0.43 | 0.99 | 0.71 | 0.98 | 0.53 | 1651 | 6077 | |||
| soybean | 1.00 | 0.79 | 1.00 | 0.96 | 1.00 | 0.86 | 2372 | 13773 | |||
| sugar beet | 1.00 | 0.83 | 1.00 | 0.90 | 1.00 | 0.86 | 1965 | 14301 | |||
| waterhemp | 0.99 | 0.70 | 0.99 | 0.38 | 0.99 | 0.50 | 4418 | 18068 | |||
| macro avg | 0.99 | 0.84 | 0.99 | 0.86 | 0.99 | 0.85 | 39028 | 209537 | |||
| weighted avg | 1.00 | 0.88 | 1.00 | 0.87 | 1.00 | 0.87 | 39028 | 209537 | |||
Finally, when comparing the crop classification results in figure 10, all eight crop classes performed well in comparison to the weed classes. When analyzing the bar graph, MobileNetV3, EfficientNet, and ConvNeXt models were the best-performing models for blackbean and canola; Xception, MobileNetV3, and EfficientNet were the best performing models for corn; EfficientNet and ConvNeXt were the best-performing models for field pea; EfficientNet was the best-performing model for flax and lentil; MobileNetV3 and Xception were the best-performing models for soybean and sugar beet, respectively. However, Densenet was the worst-performing model of all eight crop classes. In addition to this, ConvNeXt was also the worst-performing class for soybean. Moreover, blackbean and canola were the best-performing classes of all the CNN models except Densenet, where field pea was the best-performing class. Besides this, corn was also the best-performing class for the Xception,
together with blackbean and canola. However, sugar beet was the worst-performing class for Densenet and EfficientNet. Sugar beet, lentil, and flax were the worst-performing classes for MobileNetV3; sugar beet and soybean for ConvNeXt; and flax for the Xception model.
 |
| Figure 7. Normalized confusion matrix for ConvNeXt model evaluation on set 2 data. (BB: blackbean, Can: canola, FP: field pea, HW: horseweed, PA: palmer amaranth, RW: ragweed, RRPW: redroot pigweed, Soy: soybean, Sub: sugar beet, WH: waterhemp). |
 |
| Figure 8. Box plot comparing the five deep learning models. Box plots were obtained from the overall f1-score of five CNN models for set 2 test datasets. |
Discussion
This study demonstrates the performance differences on weed and crop classification when the architectures of the CNN model were different. The weed and crop RGB images acquired from the greenhouse were used to train all five CNN models, out of which ConvNeXt (Liu et al., 2022) was inspired by a vision transformer. The training datasets were large enough to train all the convolutional layers of the CNN model. Two tests were performed; first test data set was from the same experiment as data used in model training but not used in training (set 1), and the second test data set was from different batches of weed and crop planting (set 2). The model performance was superb on the first test data, with the weighted average f1-score values ranging from 0.95 to 1.00, which is in accordance with the research findings from Espejo-Garcia et al. (2020a), Hu et al. (2020), Jiang et al. (2020), (Olsen et al. (2019), and Subeesh et al. (2022). The weighted average is preferred over the macro on imbalance datasets. The values can be different for both macro and weighted average, which is clearly shown in the five tables in the results section for all five CNN models. However, model performance declined when all the models were tested with second 209,537 weed and crop test images, which is in correspondence to our previous study (GC et al., 2022) findings based on a weed classification study on different background conditions in the image. Hence, further discussion will be made based on results from the second test data because eventually models need to work on weed and crop images from different planting times and environments.
The order of model performance for the second set of test images is not exactly the same as the first set of test images. All five CNN model performances in terms of weighted average f1-score were between 0.78 and 0.88. Xception and EfficientNet outperformed the remaining three models. The Densenet model was the worst-performing model. The remaining two models’ performance was also near the level of the best-performing models, with a f1-score value of 0.87. This performance results pattern does not follow the performance pattern results obtained by Chen et al. (2022) in 15 weed classes, where Densenet was outperforming MobileNetV3, Xception, and EfficientNet (Chen et al., 2022). This makes complete sense because they have used 15 weed species without any crop species, and only three weed species (waterhemp, palmer amaranth, ragweed) used in their study are the same as our study, because of which the result can be different. However, our results correspond to the results from Espejo-Garcia et al. (2020a) in terms of the best-performing model where the Xception model, outperformed Densenet, VGG19, and InceptionResNet (Espejo-Garcia et al., 2020a).
 |
| Figure 9. Bar plot comparing the performance of five deep learning models for identifying six weed species. |
 |
| Figure 10. Bar plot comparing the performance of five deep learning models for identifying eight crop species. |
The next assessment of this study is to find the performance of individual weed and crop classes to assess the model in terms of class perspective rather than only CNN architectures prospective. This is because the high number of studies is motivated by CNN architectures prospective in their research (Ahmad et al., 2021; Espejo-Garcia et al., 2020b; Hasan et al., 2021; Subeesh et al., 2022). In our study, there was not much difference in classification performance for crop species. However, there was a range of classification performance for individual weed species, which can be clearly visualized in figure 9. The result shows that the performance of three pigweed species family weeds (palmer amaranth, redroot pigweed, and waterhemp) is low in comparison to the remaining weed species. Although these weed species belong to the same family, they have different characteristics and may need to be addressed differently to lower the weed seed bank in soil (Bensch et al., 2003; Bryson and DeFelice, 2010). The misclassification of these three species in relation to each other is high when visualizing the confusion matrix in the results section. This classification results would have improved if we considered these three weed species as pigweed class.
Furthermore, this study has used a higher number of weed and crop images to train all the layers of the CNN model. The number of weed and crop images used for training was 221,116, and the number of test images from different experiments was 209,537, which itself is a large dataset. In comparison to our study, a large number of studies on weed classification were performed with small datasets (<3200) with transfer learning approaches (Weiss et al., 2016) except a study from Trong et al. (2020), where only a few outer layers of the neural network were trained (Chen et al., 2022; Hasan et al., 2021; Lu and Young, 2020). This makes our study different from the previous studies in weed detection (Chen et al., 2022; Hasan et al., 2021; Olsen et al., 2019; Subeesh et al., 2022) in terms of image number, weed and crop species, and CNN architectures. This study was performed in greenhouse conditions, because of which it was possible to have sufficient numbers of training and testing images covering everyday images for weed and crop plants. This study does not directly contribute to the development of an artificial intelligence-based weed control system. However, the results could be useful for future researchers who want to know the results of weed classification for specific crops and weed species. Additionally, the models developed in the study could be used in object detection as backbone feature extraction layers with CNN object detection architectures.
This study might have limitations when considering the dynamic field environment. However, models trained with large numbers of weed and crop images obtained from a greenhouse can be used to build the small number of datasets in a field environment using a transfer learning approach. The better CNN model performance can be achieved when using models trained on large weed and crop image datasets rather than the different nature of ImageNet data sets for transfer learning (Weiss et al., 2016). In addition to this, sometimes it may be difficult to get all the weed species from the field to build a comprehensive weed detection model. In such scenarios, images from the greenhouse could be used to reduce the class imbalance in field data.
Conclusion
This study was successful in establishing the first single, comprehensive, multi-class weed and crop detection model for five CNN architectures. In addition, this study was based on six weed species and eight crop production systems in North Dakota, U.S., training all the millions of parameters in CNN layers with a sufficient number of images. EfficientNet and Xception proved to be the best-performing models, with a weighted average f1-score value of 0.88. These test results are from the test images obtained from different experiment sets rather than the images from the same experiment sets used to build a model. This shows the great potential to have a single, comprehensive model for weed detection in multiple crop production systems. Furthermore, this study suggests that weed or crop species with similar phenotypic traits could have negative impacts on model performance, which was clearly illustrated in this study finding for three pigweed family weeds. Another finding of this study was that the vision transformer-based ConvNeXt model was unable to outperform the Xception and EfficeintNet CNN-based models. However, the ConvNeXt model did not have a high performance difference from the EfficientNet and Xception models. This study could be the steppingstone towards the development of a single comprehensive model for eight crop production systems in North Dakota and could be beneficial for digital agriculture research in agriculture. The low performance of pigweed family weed species requires extensive study in the future to improve the classification ability of CNN models. Besides this, the weed detection model based on this study can be used as a pretrained model to build better field-based models.
Conflict of Interest
The authors declared that they have no conflicts of interest to this research. The authors also confirm that the research article titled “Multi-species Weed and Crop Classification Comparison Using Five Different Deep Learning Network Architectures” is an original work that has not been published elsewhere, nor is currently under review consideration for publication elsewhere.
Acknowledgment
This material is based upon work partially supported by the U.S. Department of Agriculture, agreement number 58-6064-8-023. Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the view of the U.S. Department of Agriculture. This work is/was supported by the USDA National Institute of Food and Agriculture, Hatch project number ND01487.
References
Ahmad, A., Saraswat, D., Aggarwal, V., Etienne, A., & Hancock, B. (2021). Performance of deep learning models for classifying and detecting common weeds in corn and soybean production systems. Comput. Electron. Agric., 184, 106081. https://doi.org/10.1016/j.compag.2021.106081
Alzubaidi, L., Zhang, J., Humaidi, A. J., Al-Dujaili, A., Duan, Y., Al-Shamma, O.,... Farhan, L. (2021). Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J. Big Data, 8(1), 53. https://doi.org/10.1186/s40537-021-00444-8
Bensch, C. N., Horak, M. J., & Peterson, D. (2003). Interference of redroot pigweed (Amaranthus retroflexus), Palmer amaranth (A. palmeri), and common waterhemp (A. rudis) in soybean. Weed Sci., 51(1), 37-43. https://doi.org/10.1614/0043-1745(2003)051[0037:IORPAR]2.0.CO;2
Bryson, C. T., & DeFelice, M. S. (2010). Weeds of the midwestern United States and central Canada. Athens, GA: University of Georgia Press.
Cech, R. M., Jovanovic, S., Kegley, S., Hertoge, K., Leisch, F., & Zaller, J. G. (2022). Reducing overall herbicide use may reduce risks to humans but increase toxic loads to honeybees, earthworms and birds. Environ. Sci. Eur., 34(1), 44. https://doi.org/10.1186/s12302-022-00622-2
Chen, D., Lu, Y., Li, Z., & Young, S. (2022). Performance evaluation of deep transfer learning on multi-class identification of common weed species in cotton production systems. Comput. Electron. Agric., 198, 107091. https://doi.org/10.1016/j.compag.2022.107091
Chollet, F. (2016). Xception: Deep learning with depthwise separable convolutions. Proc. 2017 IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (pp. 1800-1807). IEEE. https://doi.org/10.1109/CVPR.2017.195
Espejo-Garcia, B., Mylonas, N., Athanasakos, L., & Fountas, S. (2020a). Improving weeds identification with a repository of agricultural pre-trained deep neural networks. Comput. Electron. Agric., 175, 105593. https://doi.org/10.1016/j.compag.2020.105593
Espejo-Garcia, B., Mylonas, N., Athanasakos, L., Fountas, S., & Vasilakoglou, I. (2020b). Towards weeds identification assistance through transfer learning. Comput. Electron. Agric., 171, 105306. https://doi.org/10.1016/j.compag.2020.105306
GC, S., Koparan, C., Ahmed, M. R., Zhang, Y., Howatt, K., & Sun, X. (2022). A study on deep learning algorithm performance on weed and crop species identification under different image background. Artif. Intell. Agric., 6, 242-256. https://doi.org/10.1016/j.aiia.2022.11.001
Hasan, A. S., Sohel, F., Diepeveen, D., Laga, H., & Jones, M. G. (2021). A survey of deep learning techniques for weed detection from images. Comput. Electron. Agric., 184, 106067. https://doi.org/10.1016/j.compag.2021.106067
Howard, A., Sandler, M., Chen, B., Wang, W., Chen, L. C., Tan, M.,... Le, Q. (2019). Searching for MobileNetV3. Proc. 2019 IEEE/CVF Int. Conf. on Computer Vision (ICCV), (pp. 1314-1324). https://doi.org/10.1109/ICCV.2019.00140
Hu, K., Coleman, G., Zeng, S., Wang, Z., & Walsh, M. (2020). Graph weeds net: A graph-based deep learning method for weed recognition. Comput. Electron. Agric., 174, 105520. https://doi.org/10.1016/j.compag.2020.105520
Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2016). Densely connected convolutional networks. Proc. 2017 IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (pp. 2261-2269). IEEE. https://doi.org/10.1109/CVPR.2017.243
Jiang, H., Zhang, C., Qiao, Y., Zhang, Z., Zhang, W., & Song, C. (2020). CNN feature based graph convolutional network for weed and crop recognition in smart farming. Comput. Electron. Agric., 174, 105450. https://doi.org/10.1016/j.compag.2020.105450
Kamilaris, A., & Prenafeta-Boldú, F. X. (2018). Deep learning in agriculture: A survey. Comput. Electron. Agric., 147, 70-90. https://doi.org/10.1016/j.compag.2018.02.016
Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. https://doi.org/10.48550/arXiv.1412.6980
Liu, Z., Mao, H., Wu, C. Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022). A ConvNet for the 2020s. Proc. 2022 IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (pp. 11966-11976). IEEE. https://doi.org/10.1109/CVPR52688.2022.01167
Lu, Y., & Young, S. (2020). A survey of public datasets for computer vision tasks in precision agriculture. Comput. Electron. Agric., 178, 105760. https://doi.org/10.1016/j.compag.2020.105760
Oliveira, M. C., Jhala, A. J., Bernards, M. L., Proctor, C. A., Stepanovic, S., & Werle, R. (2022). Palmer Amaranth (Amaranthus palmeri) Adaptation to US Midwest Agroecosystems. Frontiers in Agronomy, 4. https://doi.org/10.3389/fagro.2022.887629
Olsen, A., Konovalov, D. A., Philippa, B., Ridd, P., Wood, J. C., Johns, J.,... White, R. D. (2019). DeepWeeds: A multiclass weed species image dataset for deep learning. Sci. Rep., 9(1), 2058. https://doi.org/10.1038/s41598-018-38343-3
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,... Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. J. Mach. Learn. Res., 12, 2825-2830.
Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. Proc. 3rd Int. Conf. on Learning Representations ICLR 2015.
Subeesh, A., Bhole, S., Singh, K., Chandel, N. S., Rajwade, Y. A., Rao, K. V.,... Jat, D. (2022). Deep convolutional neural network models for weed detection in polyhouse grown bell peppers. Artif. Intell. Agric., 6, 47-54. https://doi.org/10.1016/j.aiia.2022.01.002
Tan, M., & Le, Q. (2019). EfficientNet: Rethinking model scaling for convolutional neural networks. In C. Kamalika, & S. Ruslan (Eds.), Proceedings of the 36th International Conference on Machine Learning (Vol. 97, pp. 6105-6114). PMLR. Retrieved from https://proceedings.mlr.press/v97/tan19a.html
Thompson, J. F., Stafford, J. V., & Miller, P. C. (1991). Potential for automatic weed detection and selective herbicide application. Crop Prot., 10(4), 254-259. https://doi.org/10.1016/0261-2194(91)90002-9
Trong, V. H., Gwang-hyun, Y., Vu, D. T., & Jin-young, K. (2020). Late fusion of multimodal deep neural networks for weeds classification. Comput. Electron. Agric., 175, 105506. https://doi.org/10.1016/j.compag.2020.105506
Tudi, M., Ruan, H. D., Wang, L., Lyu, J., Sadler, R., Connell, D.,... Phung, D. T. (2021). Agriculture development, pesticide application and its impact on the environment. Int. J. Environ. Res. Public. Health, 18(3), 1112. https://doi.org/10.3390/ijerph18031112
Weiss, K., Khoshgoftaar, T. M., & Wang, D. (2016). A survey of transfer learning. J. Big Data, 3(1), 9. https://doi.org/10.1186/s40537-016-0043-6
Wu, Z., Chen, Y., Zhao, B., Kang, X., & Ding, Y. (2021). Review of weed detection methods based on computer vision. Sensors, 21(11), 3647. https://doi.org/10.3390/s21113647
Yu, J., Sharpe, S. M., Schumann, A. W., & Boyd, N. S. (2019). Deep learning for image-based weed detection in turfgrass. Eur. J. Agron., 104, 78-84. https://doi.org/10.1016/j.eja.2019.01.004