Evaluation of CNN optimizers and meta-learning approaches for classifying toxic weeds in UAV imageryABSTRACT
Brazil, the world's largest beef exporter, faces challenges in livestock farming due to the presence of noxious weeds in Brachiaria spp. pastures. The use of drones has proven to be an efficient alternative for monitoring these plants, but studies on the automatic identification of these species in aerial images are lacking. This study proposes using convolutional neural networks (CNNs) and model-agnostic meta-learning (MAML) to classify noxious weeds from drone images. Image data were collected in the municipalities of São Miguel dos Campos and Tanque D’Arca (in the state of Alagoas), using data augmentation and embedding extraction techniques to optimize model training. Several optimizers were evaluated in a CNN, with particle swarm optimization (PSO) achieving the highest accuracy (57.7 %) and the shortest processing time, followed by the Bat algorithm (BA) at 55.1 %. In the same way, MAML was tested with the 1-shot performing satisfactorily (53.3 %), while the 6-shot obtained the worst result (31.2 %). The statistical analysis confirmed significant differences among the methods, with PSO emerging as the most effective. Applying machine learning (ML) techniques to identify noxious weeds can contribute to more efficient and sustainable pasture management, reduce pesticide use, and optimize livestock productivity.
Keywords:
CNN; pasture; remote sensing; vant
Introduction
Brazil, the main exporter of bovine meat, stands out for its extensive use of pastures, mostly dominated by plants of the genus Brachiaria spp. (Malafaia et al., 2021). However, biotic stress from weeds poses a major threat to the productivity of these pastures (Mendes-Bonato et al., 2006).
The direct economic impact of weeds on the agricultural sector refers to the toxicity of certain species that can affect animal metabolism and be lethal. Some have nephrotoxic components and thorns that hinder manual removal, potentially damaging animals’ kidneys (Martins et al., 2022).
Given this scenario, precision agriculture, combined with artificial intelligence (AI) techniques and aerial imagery, has emerged as an innovative solution for mapping and monitoring these plants at an early stage, enabling prompt control (García et al., 2020; Cañón et al., 2023).
Aerial imagery can be carried out by remotely piloted aircraft, a subcategory of unmanned aerial vehicles also known as drones. These offer significant advantages over traditional methods, such as satellites, when used in precision agriculture (del Cerro et al., 2021; Gonzalez et al., 2018).
To minimize the need for human eyes to detect detailed information about agricultural stress, new AI-based technologies for drone imagery have emerged. The gamut of AI techniques includes applications such as machine learning (ML), deep learning (DL), and model-agnostic meta-learning (MAML), which have been used to identify patterns that enable the identification of weed species toxic to animals (García et al., 2020).
Different studies have integrated convolutional neural networks (CNNs) with meta-learning techniques to overcome the limitations of limited data (Wang et al., 2022b; Yao et al., 2021; Luo et al., 2022; Porto et al., 2023; Du et al., 2023) in weed detection, pest control, and the analysis of images captured by drones (Hospedales et al., 2020; Burks et al., 2005; Jiang et al., 2020; Haq, 2022).
The main objective of this work was to verify the potential of CNN-based algorithms with different optimizers and MAML for classifying weeds, pasture and toxic plants from drone imagery.
Materials and Methods
This section will address data collection and equipment, describe the types of data used during pre-processing and standardization, and finally present the methodology and the metrics used to evaluate the algorithms. The repository containing the methods and data is available at: https://github.com/joaoagro2017/mestrado.
The study was carried out in the municipalities of São Miguel dos Campos (09°46’52" S, 36°05’37" W, altitude 200 m) and Tanque D’arca (09°31’55" S, 36°25’58" W, altitude 12 m), in the state of Alagoas, in northeastern Brazil, in two experimental areas, each 4 ha in area used for cattle raising, with pasture plantations belonging to the genus Brachiaria spp. DJI Phantom 4 Advanced drone models were used for imaging, and Google Earth Pro (version 7.3.6.10441) was used to assist with the geolocation and to predefine areas to be flown over. The software programs DJI Go (version 4.3.64) and Drone Deploy (version 5.48.1) were used to pre-configure the flyovers, with information on altitude (60 m), take-off and landing points, vertical (75 %) and longitudinal (85 %) overlap rates. Points were created by the Qgis software (version 3.40.12) by identifying weed species on the orthomosaic for subsequent samples to be cut and collected to be used in the algorithm training.
A standardization method was used to pre-process drone images to ensure they had the same dimensions. Next, they were separated into classes and placed in folders labeled Weeds (class 1), Toxic Plants (class 2) and Pasture (class 3).
Data augmentation techniques such as rotations, displacements, shearing, and horizontal inversions were applied to expand the sample set (Figure 1). These techniques also fill gaps in the processing, improve the model's robustness, and minimize the risk of overfitting, especially given the reduced number of toxic weed samples identified as follows: Crotalaria retusa L. (snake rattle), Sida spp. (fanpetals or sidas), Ricinus communis L. (castor bean), Lantana camara L., and Crotalaria lanceolata E. Mey.
A) Original pasture sample and B, C and D) its transformations generated by data augmentation.
The original dataset (Table 1) presented a large sample imbalance: the Weed class had 108, the Pastures class had 42 and Toxic Plants class 35. This makes it difficult to train the model to identify patterns in minority classes, as the majority classes tend to dominate the optimization process, leading to oversampling.
Embeddings of high-dimensional vectors were generated to capture the essential features of the image data. In a vector space, semantic similarity in the images was reflected by the proximity of corresponding vectors (Shafie, 2023). The InceptionV3 model, previously trained on the ImageNet dataset, was used to generate embeddings, increasing its efficiency sufficiently for image classification (Szegedy et al., 2015).
The classification of toxic weeds was developed using a CNN with different optimizers, to learn complex, nonlinear representations of the input data.
The CNN structure contains: a) input layer; b) hidden layers; c) three dense layers with 256, 128 and 64 neurons; d) rectified linear unit; e) L2 regularization; e) batch normalization; f) dropout with 50 % rate after each layer of batch normalization; g) output layer; and h) softmax activation function.
The CNN model training applied: a) optimizer: 0.01 learning rate; b) loss function; and c) regularization: early stopping.
Optimizers were used to adjust the network parameters. The adaptive moment estimation (ADAM) optimizer is efficient for deep neural network problems. According to Kingma and Ba (2014), ADAM combines the advantages of two other optimization methods: stochastic gradient with momentum and RMSProp regularization. It individually adjusts the learning rate of each parameter and achieves rapid convergence, even in large deep neural networks (Wang et al., 2022b). This optimizer was configured with a learning rate of 0.0001, which favors gradual convergence and helps avoid large oscillations in weight updates.
Adaptive moment estimation with decoupled weight decay (ADAMW) is an optimizer variant of ADAM (Kingma and Ba, 2014), that addresses a limitation of L2 regularization by decoupling weight decay from gradient updates, thereby improving generalization across multiple DL tasks.
This allows ADAMW to achieve more efficient parameter learning, resulting in superior performance compared to traditional ADAM, especially in image classification and other complex tasks. It offers a robust solution for optimizing deep neural networks, resulting in better performance and generalization than conventional L2 regularization (Kingma and Ba, 2014).
Particle swarm optimization (PSO) is an optimization algorithm inspired by the collective behavior of birds searching for food. It uses particles that "fly" in a multidimensional search space, each with an updated position and velocity with each iteration following the best individual position and the best overall position of all particles (Kamalova et al., 2017; Wang et al., 2022b)
The Bat algorithm (BA) is an optimizer for swarm intelligence that simulates the echolocation bats’ behavior for food location (Yang, 2010). This algorithm proposes a time factor improvement over the early BA, focusing on the search range in the initial stage and mining capacity in the next stage. (Wang et al., 2022b).
Model-agnostic meta-learning aims to solve generalization and speed of learning problems in scenarios with few samples, and its training process involves two main cycles: 1) inner loop: allows the model to find optimal parameters for each particular task. (Du et al., 2023; Wang et al., 2022b); 2) outer loop: seeks a set of initial parameters sufficiently generalist to accelerate learning in new tasks with few data. (Du et al., 2023; Wang et al., 2022b)
This work fused MAML with a self-constructed CNN, which, with ideal initial parameters, accelerates training and improves model generalization. CNN consists of nine layers: four convolutions, four pooling layers, and one fully connected layer. Three comparisons were made with the MAML using the following few-shot learning techniques: 1-shot, 2-shot, and 6-shot (Wang et al., 2022b).
K-fold cross validation was used to evaluate the methods by dividing the dataset into K parts (folds). In each iteration, one of these folds is used as a test set, while the remaining K-1 folds are used for training (Figure 2). The process was repeated K times, alternating the fold used for testing until all had been used. This ensured that each sample participated in both training and testing, increasing model reliability and preventing overfitting. In this study, K = 10 was used for all algorithms.
The model's classification performance was evaluated using Friedman test, a non-parametric method for comparing treatments with repeated measurements. The accuracy results were obtained for each cross-validation fold, using different optimizers. The Friedman test was sufficiently accurate for identifying differences between algorithms across folds, without the need to assume normality in the data (Demšar, 2006).
The accuracy values for each fold were organized in a matrix, with rows representing the folds and columns representing the evaluated models. After the test application, a p value of < 0.05 was obtained, indicating rejection of the null hypothesis because all models showed statistically equivalent performance.
After this, it was necessary to identify which pairs of models differed in performance. Next, a Conover post-test was performed to adjust p-values using procedures such as Holm-Bonferroni and to minimize the risk of type I errors in multiple comparisons (Conover, 1999).
The evaluation metrics’ accuracy, precision, recall, and F1-score were used to evaluate the CNN performance for image classification. These metrics indicate the model's accuracy in correctly predicting image classes.
The accuracy metric calculates the ratio of correctly classified images to the total number of images in the dataset and evaluates the overall model accuracy across all classes (Eq. 1).
Precision metric measures the ratio of true positives (TP) to the sum of true and false positives (FP) (Eq. 2).
Recall is the ratio of TP to the sum of TP and false negatives (FN) (Eq. 3).
The F1-score metric is the harmonic mean of accuracy and recall, providing a single metric that accounts for both FP and FN (Eq. 4).
Results
Training and validation of the drone images were carried out with 100 epochs, evaluating each metric of accuracy and computational processing time for each optimizer (Table 2). PSO optimizer achieved the best metrics among the others. It had the highest values for accuracy (57.70 %), precision (57.30 %), recall (57.70 %), and F1-score (57.00 %), in addition to having the lowest processing time (101.15 s). PSO enabled provided the identification of patterns even from a small number of samples (Kamalova et al., 2017).
The BA recorded values similar to PSO in terms of accuracy (55.10 %), precision (55.20 %), recall (55.10 %), and F1-score (55.00 %), which is a good balance between accuracy and processing time, as well as promising as regards computational efficiency (Yang, 2010).
These results showed that PSO is the most advantageous choice for these types of image data, while BA is a viable alternative in resource-limited scenarios, which corroborates previous studies with these optimizers (Kamalova et al., 2017; Yang, 2010).
Adaptive moment estimation and ADAMW had the worst performance, with results similar to each other at 53.80 and 54.20 %, respectively. Processing times were also worse, especially for ADAMW, with a speed limitation due to applicability (Wang et al., 2022b). ADAM had the lowest overall performance, with an accuracy of 53.80 % and a time of 102.68 s. Although reliable for several applications, its performance was inferior to others in this experiment (Wang et al., 2022b).
The values for training and validation were obtained after 100 epochs with MAML (Table 3). The 1-shot had the best results in terms of accuracy (53.3 %), precision (53.0 %), recall (53.3 %), F1-score (52.8 %), and runtime 96.98 s. Therefore, increasing the number of shots did not generate any relevant improvement (Du et al., 2023; Wang et al., 2022b).
The 2-shot model underperformed the 1-shot in accuracy (45.0 %), precision (44.7 %), precision (45.0 %) F1-score (44.7 %), and processing time (97.99 s). The 6-shot had the worst performance among the models in accuracy (31.2 %), precision (30.9 %), recall (31.2 %), F1-score (30.4 %), and execution time (105.07 s). These results indicate that increasing the number of shots is not advantageous for performance (Du et al., 2023).
As for the differences in the performance of the models, the Friedman test indicated a relevant discrepancy between the models due to the p-value of 3.18 × 10−6.
The p-value served as the reference for comparisons with results from the Conover test, adjusting the p-values to control the type I error. The p-value matrix is shown in Table 4, where each entry corresponds to the comparison between pairs of models.
According to the p-values obtained here, the MAML 2-shot and 6-shot algorithms showed the greatest differences (Table 4), with lower values than those of other methods (Conover, 1999). The PSO algorithm also stood out with lower values than ADAM (0.069408) and ADAMW (0.066084), suggesting statistical differences. Nevertheless, ADAM and ADAMW had a high value (0.9816), indicating no statistical difference between them. When comparing BA with ADAM (0.7940) and MAML 1-shot (0.5200), no differences were observed between them. Thus, the greatest contrast between the methods was between MAML 2-shot and 6-shot other algorithms (Conover, 1999).
A critical difference diagram (Figure 3) was generated from the Conover test values to illustrate the differences in the average performance ranks of the models. The diagram highlights the groups of models that showed no statistical differences, with connections represented by the black horizontal lines.
Critical difference diagram. MAML = model-agnostic meta-learning; CNN = convolutional neural network; PSO = particle swarm optimization; ADAM = adaptive moment estimation; ADAMW = ADAM with decoupled weight decay; BA = bat algorithm.
The MAML 6-shot and 2-shot models achieved average performance, ranking lowest, whereas PSO ranked highest. This diagram facilitates identifying methods that differ in performance, interpreting the results in Table 4 and choosing the most appropriate model.
Results for algorithms’ processing time showed different performance and consistency (Table 5). ADAM showed consistent performance, with a low percentage difference (–0.39 %), indicating great efficiency, but had the worst time with the greatest positive difference (9.19 %). PSO showed a small difference (–1.87 %), reflecting overall time down. BA had a balanced performance, but with a positive difference (1.89 %), slightly above average. MAML 1-shot and 2-shot showed greater differences from the overall average –5.92 and –4.94 %, respectively, as well as the best computational processing times. However, the MAML 6-shot showed a positive difference (2.03 %), and the second worst time.
Discussion
MAML's results on scarce data showed clear limitations in the analyzed context. In scenarios with greater data availability, traditional optimization methods, such as PSO and BA, presented more robust results, with greater accuracy and better balance between performance and processing time (Li and Yang, 2021). On the other hand, MAML can be a valid alternative in terms of reduced time and costs. Meta-learning efficiency for classification and species identification with few samples should be explored for future optimization, although computational cost and adjustments by the models reinforce the need for careful evaluation in such a context (Du et al., 2023; Wang et al., 2022a; Wang et al., 2022b).
The MAML approach with few-shot learning suggests that, while meta-learning can be promising for some applications, its computational cost and the need for more shots may not always be advantageous, especially in situations with few data and specific characteristics (Li and Yang, 2021). However, the PSO-optimized model achieved the highest accuracy in this context for identifying toxic weeds using drone imagery.
The Friedman test indicated that the statistical analysis revealed significant differences in the performance of the analyzed algorithms.
The performance of the resultant Conover post-test indicated differences between the algorithms and the others, and the adjustment in the p-value ensured control of type I errors.
The results revealed that meta-learning-based models, such as MAML 6-shot and MAML 2-shot, had the lowest average rankings compared to the other models. The MAML 6-shot presented negative statistical differences among the other models.
On the other hand, ADAM, ADAMW, BA, and MAML 1-shot methods did not show significant differences among themselves (p > 0.05). Despite this, CNN and PSO had the highest average rank (0.91), while MAML 6-shot the lowest (0.19).
The growth of CNN technology for plant identification represents a revolution in agriculture. CNNs have a recognition capacity similar to that of human brain but can assimilate data faster (Jati and Georgiou, 2019). It is possible to distinguish toxic weed species of Fabaceae family in the grassland genus of Brachiaria spp. by drone image, even when the number of samples is limited (Mesías-Ruiz et al., 2024).
When choosing an appropriate algorithm, it must be selected based on the specific context and available data. This is especially important when dealing with drone imagery from 60 m altitudes, since the type of data directly influences the model's performance (Zhao et al., 2024). The identification of toxic plants is fundamental to animal health and plant management. Accidental ingestion of toxic weeds can lead to serious consequences such as intoxication, organ impairment, and even death (Martins et al., 2022).
Detection technologies using CNNs and MAML can facilitate pasture monitoring. Drone images combined with AI algorithms allow for the identification of toxic plants, enabling rapid management interventions, minimizing animal intoxication, and promoting their food security (Kaya and Goraj, 2020).
Morphological similarities between certain plant species can be difficult to identify (Yang et al., 2022). However, the differentiation between monocots and dicots was effective in the study, demonstrating the potential of CNNs to identify pastures infested with Fabaceae species known to be toxic (Mesías-Ruiz et al., 2024).
Excessive use of epochs during CNN training can lead to overfitting, compromising the algorithm's accuracy (Ying, 2019).
These results emphasize the potential of toxic weed identification and classification using aerial images and DL techniques in precision agriculture. The adoption of such advanced technologies can provide outstanding monitoring solutions to the challenges faced in livestock farming, especially in pasture management.
The results showed that a CNN-based algorithm using optimizers was suitable for the recognition of toxic weed species in drone images at an altitude of 60 m.
The importance of technologies in integrating agricultural practices and ML into toxic weed detection is directly linked to the management of animal grazing to ensure their welfare. The use of DL methods such as CNNs and MAML can help monitor crops imaged by drones, enabling fast, precise control. This is crucial to contemporary farming, because of the limited data for identifying these toxic species.
According to the statistical analysis, MAML 6-shot had the lowest average ranking, that is, the worst results in the tests. On the other hand, the PSO optimizer obtained the highest average ranking, considered the best results. The MAML 1-shot model achieved accuracy levels similar to PSO, with lower processing time than the other models, indicating that even though it did not achieve the highest statistical accuracy, it did not differ from other methods, that recorded better results.
This study represents a scientific advance by proposing a new approach to identifying toxic plants in pasture areas, that combines CNNs with meta-learning techniques. Unlike existing CNN approaches that focus on plant species classification, these combinations of techniques require less data and yield classification results more quickly and with less computational capacity. Thus, this study not only fills a gap in the automated mapping of these species via drone imagery, but also demonstrates the feasibility of few-example learning methods for agricultural applications, thereby contributing to reductions in animal losses due to poisoning.
Therefore, classification technologies by CNN algorithms applied in drone imagery for pasture areas represents a breakthrough for the livestock sector in Alagoas. Accurate detection and control of toxic weeds enable targeted pesticide applications, reducing the use of chemical products and their environmental impact. In addition, this solution promotes more efficient and sustainable agricultural production, thereby contributing to environmental preservation and increasing in productivity.
Data availability statement
The authors declare that data supporting the results of this study are available within the article and its supplementary materials.
Acknowledgments
We would like to express our sincere thanks to the Yathe Agritec startup for supplying the data used in this study, which was essential to carrying out this work. We would also like to thank the Universidade Federal de Alogoas (UFAL), especially the Pós-graduação em Informática at the Instituto de Computação (IC/UFAL), for their partnership and continuous support in the development of this project. Without the support and collaboration of these institutions, this work would not have been possible.
References
-
Haq MA. 2022. CNN Based Automated Weed Detection System Using UAV Imagery. Computer Systems Science and Engineering 42: 837-849. https://doi.org/10.32604/csse.2022.023016
» https://doi.org/10.32604/csse.2022.023016 -
Mendes-Bonato AB, Risso-Pascotto C, Pagliarini MS, Do Valle CB. 2006. Chromosome number and meiotic behaviour in Brachiaria jubata (Gramineae). Journal of Genetics 85: 83-87. https://doi.org/10.1007/BF02728976
» https://doi.org/10.1007/BF02728976 -
Burks TF, Shearer SA, Heath JR, Donohue KD. 2005. Evaluation of neural-network classifiers for weed species discrimination. Biosystems Engineering 91: 293-304. https://doi.org/10.1016/j.biosystemseng.2004.12.012
» https://doi.org/10.1016/j.biosystemseng.2004.12.012 -
Cañón LVG, Rodríguez M, Pinzón NAJ, Ortega RO, Mesa ALH. 2023. Drones in agriculture 4.0. Ingeniería Solidaria 19: 1-27 (in Spanish, with abstract in English). https://doi.org/10.16925/2357-6014.2023.02.05
» https://doi.org/10.16925/2357-6014.2023.02.05 -
del Cerro J, Ulloa CC, Barrientos A, Rivas JL. 2021. Unmanned aerial vehicles in agriculture: a survey. Agronomy 11: 203. https://doi.org/10.3390/agronomy11020203
» https://doi.org/10.3390/agronomy11020203 - Conover WJ. 1999. Practical Nonparametric Statistics. 3 ed. Wiley, New York, NY, USA.
- Demšar J. 2006. Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research 7: 1-30.
-
Du M, Nair S, Sadigh D, Finn C. 2023. Behavior retrieval: few-shot imitation learning by querying unlabeled datasets. arXiv: 2304.08742. https://doi.org/10.48550/arXiv.2304.08742
» https://doi.org/10.48550/arXiv.2304.08742 -
García R, Aguilar J, Toro M, Pinto A, Rodríguez P. 2020. A systematic literature review on the use of machine learning in precision livestock farming. Computers and Electronics in Agriculture 179: 105826. https://doi.org/10.1016/j.compag.2020.105826
» https://doi.org/10.1016/j.compag.2020.105826 -
Gonzalez F, Mcfadyen A, Puig E. 2018. Advances in Unmanned Aerial Systems and Payload Technologies for Precision Agriculture. p. 133-155. In: Chen G. ed. Advances in Agricultural Machinery and Technologies. CRC Press, Boca Raton, FL, USA. https://doi.org/10.1201/9781351132398-6
» https://doi.org/10.1201/9781351132398-6 -
Hospedales T, Antoniou A, Micaelli P, Storkey A. 2020. Meta-learning in neural networks: a survey. arXiv: 2004.05439. https://doi.org/10.1109/TPAMI.2021.3079209
» https://doi.org/10.1109/TPAMI.2021.3079209 -
Jati A, Georgiou P. 2019. Neural predictive coding using convolutional neural networks toward unsupervised learning of speaker characteristics. IEEE/ACM Transactions on Audio, Speech, and Language Processing 27: 1577-1589. https://doi.org/10.1109/TASLP.2019.2921890
» https://doi.org/10.1109/TASLP.2019.2921890 -
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. Computers and Electronics in Agriculture 174: 105450. https://doi.org/10.1016/j.compag.2020.105450
» https://doi.org/10.1016/j.compag.2020.105450 -
Kamalova DI, Galimullin DZ, Sibgatullin ME, Salakhov MK. 2017. An evolutionary particle swarm optimization algorithm for mathematical processing of experimental spectra. Optics and Spectroscopy 122: 687-691. https://doi.org/10.1134/S0030400X17050101
» https://doi.org/10.1134/S0030400X17050101 -
Kaya S, Goraj Z. 2020. The use of drones in agricultural production. International Journal of Innovative Approaches in Agricultural Research 4: 166-176. https://doi.org/10.29329/ijiaar.2020.254.2
» https://doi.org/10.29329/ijiaar.2020.254.2 -
Kingma DP, Ba J. 2014. Adam: a method for stochastic optimization. arXiv: 1412.6980. https://doi.org/10.48550/arXiv.1412.6980
» https://doi.org/10.48550/arXiv.1412.6980 -
Li Y, Yang J. 2021. Meta-learning baselines and database for few-shot classification in agriculture. Computers and Electronics in Agriculture 182: 106055. https://doi.org/10.1016/j.compag.2021.105055
» https://doi.org/10.1016/j.compag.2021.105055 -
Luo S, Li Y, Gao P, Wang Y, Serikawa S. 2022. Meta-seg: A survey of meta-learning for image segmentation. Pattern Recognition 126: 108586. https://doi.org/10.1016/j.patcog.2022.108586
» https://doi.org/10.1016/j.patcog.2022.108586 -
Malafaia GC, Mores GV, Casagranda YG, Barcellos JOJ, Costa FP. 2021. The Brazilian beef cattle supply chain in the next decades. Livestock Science 253: 104704. https://doi.org/10.1016/j.livsci.2021.104704
» https://doi.org/10.1016/j.livsci.2021.104704 -
Martins D, Marchi SR, Marques RF. 2022. Pasture weed biology and management in Brazil. Semina - Ciências Agrárias 43: 459-482. https://doi.org/10.5433/1679-0359.2022v43n1p459
» https://doi.org/10.5433/1679-0359.2022v43n1p459 -
Mesías-Ruiz GA, Borra-Serrano I, Peña JM, Castro AI, Fernández-Quintanilla C, Dorado J. 2024. Weed species classification with UAV imagery and standard CNN models: Assessing the frontiers of training and inference phases. Crop Protection 182: 106721. https://doi.org/10.1016/j.cropro.2024.106721
» https://doi.org/10.1016/j.cropro.2024.106721 -
Porto JVA, Dorsa AC, Weber VAM, Porto KRA, Pistori H. 2023. Usage of few-shot learning and meta-learning in agriculture: A literature review. Smart Agricultural Technology 5: 100307. https://doi.org/10.1016/j.atech.2023.100307
» https://doi.org/10.1016/j.atech.2023.100307 -
Shafie E. 2023. An empirical study of extracting embedded text from digital images. International Journal of Advanced and Applied Sciences 10: 48-53. https://doi.org/10.21833/ijaas.2023.06.006
» https://doi.org/10.21833/ijaas.2023.06.006 -
Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. 2015. Rethinking the inception architecture for computer vision. arXiv: 1512.00567. https://doi.org/10.48550/arXiv.1512.00567
» https://doi.org/10.48550/arXiv.1512.00567 -
Wang D, Cao W, Zhang F, Li Z, Xu S, Wu X. 2022a. A review of deep learning in multiscale agricultural sensing. Remote Sensing 14: 559. https://doi.org/10.3390/rs14030559
» https://doi.org/10.3390/rs14030559 -
Wang S, Han Y, Chen J, He X, Zhang Z, Liu X, et al. 2022b. Weed density extraction based on few-shot learning through UAV remote sensing RGB and multispectral images in ecological irrigation area. Frontiers in Plant Science 12: 735230. https://doi.org/10.3389/fpls.2021.735230
» https://doi.org/10.3389/fpls.2021.735230 -
Yang J, Wang Y, Chen Y, Yu J. 2022. Detection of weeds growing in alfalfa using convolutional neural networks. Agronomy 12: 1459. https://doi.org/10.3390/agronomy12061459
» https://doi.org/10.3390/agronomy12061459 -
Yang X-S. 2010. A New Metaheuristic Bat-Inspired Algorithm. p. 65-74. In: González JR, Pelta DA, Cruz C, Terrazas G, Krasnogor N. eds. Nature Inspired Cooperative Strategies for Optimization. Springer, Berlin, Germany. https://doi.org/10.1007/978-3-642-12538-6_6
» https://doi.org/10.1007/978-3-642-12538-6_6 -
Yao X, Zhu J, Huo G, Xu N, Liu X, Zhang C. 2021. Model-agnostic multi-stage loss optimization meta learning. International Journal of Machine Learning and Cybernetics 12: 2349-2363. https://doi.org/10.1007/s13042-021-01316-6
» https://doi.org/10.1007/s13042-021-01316-6 -
Ying X. 2019. An overview of overfitting and its solutions. Journal of Physics - Conference Series 1168: 022022. https://doi.org/10.1088/1742-6596/1168/2/022022
» https://doi.org/10.1088/1742-6596/1168/2/022022 -
Zhao D, Shao F, Liu Q, Yang L, Zhang H, Zhang Z. 2024. A small object detection method for drone-captured images based on improved YOLOv7. Remote Sensing 16: 1002. https://doi.org/10.3390/rs16061002
» https://doi.org/10.3390/rs16061002
Edited by
-
Edited by:
Thiago Libório Romanelli https://orcid.org/0000-0001-5407-9031
