Rui Mao, Yuchen Zhang, Zexi Wang, Xingan Hao, Tao Zhu, Shengchang Gao, Xiaoping Hu
{"title":"DAE-Mask: a novel deep-learning-based automatic detection model for in-field wheat diseases","authors":"Rui Mao, Yuchen Zhang, Zexi Wang, Xingan Hao, Tao Zhu, Shengchang Gao, Xiaoping Hu","doi":"10.1007/s11119-023-10093-x","DOIUrl":null,"url":null,"abstract":"<p>Wheat diseases seriously restrict the safety of wheat production and food quality. For farmers and agriculture technicians, diagnosing the disease with the naked eye is not suitable for modern precision agriculture. Deep learning has shown promise in crop disease diagnosis, but accuracy and speed remain a significant challenge in natural field conditions. In this study, a novel DAE-Mask method based on <b>d</b>iversification-<b>a</b>ugmented features and <b>e</b>dge features was proposed for intelligent wheat disease detection. DAE-Mask used Densely Connected Convolutional Networks (DenseNet) for preliminary feature extraction, and a backbone feature extraction network combining Feature Pyramid Network (FPN) and attention mechanism was designed to extract diversification-augmented features. To accelerate DAE-Mask, an Edge Agreement Head module based on Sobel filters was designed to compare edge features during training, which improved the model’s mask generation efficiency. We also built a multi-scene wheat disease dataset, MSWDD2022, containing images of wheat stripe rust, wheat powdery mildew, wheat yellow dwarf, and wheat scab. Our model achieved detection speed of 0.08s/pic. On MSWDD2022, our model with mean average precision (<i>mAP</i>) of 96.02% outperformed YOLOv5s, YOLOv8x, SSD, EfficientDet, CenterNet, and RefineDet by 7.79, 1.32, 3.54, 4.79, 9.77, and 5.29 percentage points, respectively. On the public dataset PlantDoc, our model with <i>mAP</i> of 57.68% outperformed YOLOv5s, YOLOv8x, SSD, EfficientDet, CenterNet, and RefineDet by 27.76, 6.48, 14.43, 11.79, 19.40, and 13.40 percentage points, respectively. Finally, the DAE-Mask was deployed on WeChat Mini Program to realize the real-time detection of in-field wheat diseases. The <i>mAP</i> reached 92.78%, and the average return delay of each image was 1.43s.</p>","PeriodicalId":20423,"journal":{"name":"Precision Agriculture","volume":null,"pages":null},"PeriodicalIF":5.4000,"publicationDate":"2023-12-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Precision Agriculture","FirstCategoryId":"97","ListUrlMain":"https://doi.org/10.1007/s11119-023-10093-x","RegionNum":2,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"AGRICULTURE, MULTIDISCIPLINARY","Score":null,"Total":0}
引用次数: 0
Abstract
Wheat diseases seriously restrict the safety of wheat production and food quality. For farmers and agriculture technicians, diagnosing the disease with the naked eye is not suitable for modern precision agriculture. Deep learning has shown promise in crop disease diagnosis, but accuracy and speed remain a significant challenge in natural field conditions. In this study, a novel DAE-Mask method based on diversification-augmented features and edge features was proposed for intelligent wheat disease detection. DAE-Mask used Densely Connected Convolutional Networks (DenseNet) for preliminary feature extraction, and a backbone feature extraction network combining Feature Pyramid Network (FPN) and attention mechanism was designed to extract diversification-augmented features. To accelerate DAE-Mask, an Edge Agreement Head module based on Sobel filters was designed to compare edge features during training, which improved the model’s mask generation efficiency. We also built a multi-scene wheat disease dataset, MSWDD2022, containing images of wheat stripe rust, wheat powdery mildew, wheat yellow dwarf, and wheat scab. Our model achieved detection speed of 0.08s/pic. On MSWDD2022, our model with mean average precision (mAP) of 96.02% outperformed YOLOv5s, YOLOv8x, SSD, EfficientDet, CenterNet, and RefineDet by 7.79, 1.32, 3.54, 4.79, 9.77, and 5.29 percentage points, respectively. On the public dataset PlantDoc, our model with mAP of 57.68% outperformed YOLOv5s, YOLOv8x, SSD, EfficientDet, CenterNet, and RefineDet by 27.76, 6.48, 14.43, 11.79, 19.40, and 13.40 percentage points, respectively. Finally, the DAE-Mask was deployed on WeChat Mini Program to realize the real-time detection of in-field wheat diseases. The mAP reached 92.78%, and the average return delay of each image was 1.43s.
期刊介绍:
Precision Agriculture promotes the most innovative results coming from the research in the field of precision agriculture. It provides an effective forum for disseminating original and fundamental research and experience in the rapidly advancing area of precision farming.
There are many topics in the field of precision agriculture; therefore, the topics that are addressed include, but are not limited to:
Natural Resources Variability: Soil and landscape variability, digital elevation models, soil mapping, geostatistics, geographic information systems, microclimate, weather forecasting, remote sensing, management units, scale, etc.
Managing Variability: Sampling techniques, site-specific nutrient and crop protection chemical recommendation, crop quality, tillage, seed density, seed variety, yield mapping, remote sensing, record keeping systems, data interpretation and use, crops (corn, wheat, sugar beets, potatoes, peanut, cotton, vegetables, etc.), management scale, etc.
Engineering Technology: Computers, positioning systems, DGPS, machinery, tillage, planting, nutrient and crop protection implements, manure, irrigation, fertigation, yield monitor and mapping, soil physical and chemical characteristic sensors, weed/pest mapping, etc.
Profitability: MEY, net returns, BMPs, optimum recommendations, crop quality, technology cost, sustainability, social impacts, marketing, cooperatives, farm scale, crop type, etc.
Environment: Nutrient, crop protection chemicals, sediments, leaching, runoff, practices, field, watershed, on/off farm, artificial drainage, ground water, surface water, etc.
Technology Transfer: Skill needs, education, training, outreach, methods, surveys, agri-business, producers, distance education, Internet, simulations models, decision support systems, expert systems, on-farm experimentation, partnerships, quality of rural life, etc.