Improving Accuracy of Tomato Plant Disease Diagnosis Based on Deep Learning With Explicit Control of Hidden Classes

Fuentes, Alvaro; Yoon, Sook; Lee, Mun Haeng; Park, Dong Sun

doi:10.3389/fpls.2021.682230

Cited by 19 publications

(8 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The prominent advantage of deep learning is feature learning; that is, the combination of lower-level features to form higher-level features ( LeCun et al., 2015 ). This technique can improve the accuracy of detection, classification, and recognition; for example, after several iterative trainings in disease recognition, the recognition accuracy can reach more than 90% ( Fuentes et al., 2021 ). Second, deep learning involves versatility and generalization.…”

Section: Discussionmentioning

confidence: 99%

Integrated network analyses identify MYB4R1 neofunctionalization in the UV-B adaptation of Tartary buckwheat

Liu

Sun

et al. 2022

Plant Communications

View full text Add to dashboard Cite

Section: Discussionmentioning

confidence: 99%

Integrated network analyses identify MYB4R1 neofunctionalization in the UV-B adaptation of Tartary buckwheat

Liu

Sun

et al. 2022

Plant Communications

View full text Add to dashboard Cite

“…In the literature, this problem has barely been investigated in plant disease recognition; however, it is an important issue for developing a more generalized model. Early work in this area ( Fuentes et al., 2021b ) shows the benefits of using control classes, such as background and healthy leaves, to lead the learning process toward classes of interest. It exhibited improved performance as an easier-to-adapt model across environments.…”

Section: Limited Datasetmentioning

confidence: 99%

Embracing limited and imperfect training datasets: opportunities and challenges in plant disease recognition using deep learning

Xu,

Kim,

Yang

et al. 2023

Front. Plant Sci.

Self Cite

View full text Add to dashboard Cite

Recent advancements in deep learning have brought significant improvements to plant disease recognition. However, achieving satisfactory performance often requires high-quality training datasets, which are challenging and expensive to collect. Consequently, the practical application of current deep learning–based methods in real-world scenarios is hindered by the scarcity of high-quality datasets. In this paper, we argue that embracing poor datasets is viable and aims to explicitly define the challenges associated with using these datasets. To delve into this topic, we analyze the characteristics of high-quality datasets, namely, large-scale images and desired annotation, and contrast them with the limited and imperfect nature of poor datasets. Challenges arise when the training datasets deviate from these characteristics. To provide a comprehensive understanding, we propose a novel and informative taxonomy that categorizes these challenges. Furthermore, we offer a brief overview of existing studies and approaches that address these challenges. We point out that our paper sheds light on the importance of embracing poor datasets, enhances the understanding of the associated challenges, and contributes to the ambitious objective of deploying deep learning in real-world applications. To facilitate the progress, we finally describe several outstanding questions and point out potential future directions. Although our primary focus is on plant disease recognition, we emphasize that the principles of embracing and analyzing poor datasets are applicable to a wider range of domains, including agriculture. Our project is public available at https://github.com/xml94/EmbracingLimitedImperfectTrainingDatasets.

show abstract

“…S13). This low rate of FP indicates that the model seldom misidenti es non-plant objects as plants [113,114]. Additionally, the False Negative (FN) rate of 18.8% demonstrates that the model misses a relatively small proportion of actual plants in the eld.…”

Section: Plant Detection Accuracymentioning

confidence: 99%

Digital Phenotyping in Plant Breeding: Evaluating Relative Maturity, Stand Count, and Plant Height in Dry Beans (Phaseolus vulgaris L.) via RGB Drone-Based Imagery and Deep Learning Approaches

Volpato

Wright

Gomez

2023

Preprint

View full text Add to dashboard Cite

Background Significant effort has been made in manually tracking plant maturity and to measure early-stage plant density, and crop height in experimental breeding plots. Agronomic traits such as relative maturity (RM), stand count (SC) and plant height (PH) are essential to cultivar development, production recommendations and management practices. The use of RGB images collected via drones may replace traditional measurements in field trials with improved throughput, accuracy, and reduced cost. Recent advances in deep learning (DL) approaches have enabled the development of automated high-throughput phenotyping (HTP) systems that can quickly and accurately measure target traits using low-cost RGB drones. In this study, a time series of drone images was employed to estimate dry bean relative maturity (RM) using a hybrid model combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for features extraction and capturing the sequential behavior of time series data. The performance of the Faster-RCNN object detection algorithm was also examined for stand count (SC) assessment during the early growth stages of dry beans. Various factors, such as flight frequencies, image resolution, and data augmentation, along with pseudo-labeling techniques, were investigated to enhance the performance and accuracy of DL models. Traditional methods involving pre-processing of images were also compared to the DL models employed in this study. Moreover, plant architecture was analyzed to extract plant height (PH) using digital surface model (DSM) and point cloud (PC) data sources. Results The CNN-LSTM model demonstrated high performance in predicting the RM of plots across diverse environments and flight datasets, regardless of image size or flight frequency. The DL model consistently outperformed the pre-processing images approach using traditional analysis (LOESS and SEG models), particularly when comparing errors using mean absolute error (MAE), providing less than two days of error in prediction across all environments. When growing degree days (GDD) data was incorporated into the CNN-LSTM model, the performance improved in certain environments, especially under unfavorable environmental conditions or weather stress. However, in other environments, the CNN-LSTM model performed similarly to or slightly better than the CNN-LSTM + GDD model. Consequently, incorporating GDD may not be necessary unless weather conditions are extreme. The Faster R-CNN model employed in this study was successful in accurately identifying bean plants at early growth stages, with correlations between the predicted SC and ground truth (GT) measurements of 0.8. The model performed consistently across various flight altitudes, and its accuracy was better compared to traditional segmentation methods using pre-processing images in OpenCV and the watershed algorithm. An appropriate growth stage should be carefully targeted for optimal results, as well as precise boundary box annotations. On average, the PC data source marginally outperformed the CSM/DSM data to estimating PH, with average correlation results of 0.55 for PC and 0.52 for CSM/DSM. The choice between them may depend on the specific environment and flight conditions, as the PH performance estimation is similar in the analyzed scenarios. However, the ground and vegetation elevation estimates can be optimized by deploying different thresholds and metrics to classify the data and perform the height extraction, respectively. Conclusions The results demonstrate that the CNN-LSTM and Faster R-CNN deep learning models outperforms other state-of-the-art techniques to quantify, respectively, RM and SC. The subtraction method proposed for estimating PH in the absence of accurate ground elevation data yielded results comparable to the difference-based method. In addition, open-source software developed to conduct the PH and RM analyses can contribute greatly to the phenotyping community.

show abstract

Improving Accuracy of Tomato Plant Disease Diagnosis Based on Deep Learning With Explicit Control of Hidden Classes

Cited by 19 publications

References 33 publications

Integrated network analyses identify MYB4R1 neofunctionalization in the UV-B adaptation of Tartary buckwheat

Integrated network analyses identify MYB4R1 neofunctionalization in the UV-B adaptation of Tartary buckwheat

Embracing limited and imperfect training datasets: opportunities and challenges in plant disease recognition using deep learning

Digital Phenotyping in Plant Breeding: Evaluating Relative Maturity, Stand Count, and Plant Height in Dry Beans (Phaseolus vulgaris L.) via RGB Drone-Based Imagery and Deep Learning Approaches

Contact Info

Product

Resources

About