Search

The rise of deep learning techniques has profoundly impacted both research and applications of pattern and object recognition in digital images. In plant pathology, the number of scientific articles on the use of deep learning for disease classification using images has grown steadily for at least a decade and targeted most important agricultural crops. Results have been encouraging, with accuracies of many prediction models usually approaching 100%. It is now widely accepted that, enough data being available, deep learning models can solve most of the image classification problems. However, determining what “enough” means in each context is far from trivial because this involves not only the number of samples used for training, but also the quality, in particular the representativeness of the dataset. More important than having a large sample size is to guarantee that all the variability associated to a given classification problem is represented in the dataset. Achieving this goal is particularly challenging for plant disease images because the agricultural environment is non-structured and highly dynamic, containing numerous variables that introduce variability to the problem. To make matters even more difficult, image annotation is time consuming and prone to inconsistencies due to its subjectivity. As a result, all studies in the literature employ datasets that represent only a fraction of the whole range of the variability, and many of these do not even acknowledge the limitations of the experimental conditions. Experiments with limited scope are valuable in the early stages of emerging research topics, but the application of deep learning to plant pathology has matured to the point where new studies need to contribute something more substantial. Unfortunately, many of the recent publications have been redundant, differing from previous research only by the adoption of slightly different experimental setups and improved model architectures. To move forward, new studies in this field need to address the data gap problem more effectively. This article delves deep into some technical and practical issues to achieve this goal and to increase the usefulness of the future studies. Although this article is dedicated primarily to proximal images, many of the remarks also hold for images captured using unmanned aerial vehicles.

Phytopathometry can be defined as the branch of plant pathology (phytopathology) that is concerned with estimation or measurement of the amount of plant disease expressed by symptoms of disease or signs of a pathogen on a single or group of specimens. Phytopathometry is critical for many reasons, including analyzing yield loss due to disease, breeding for disease resistance, evaluating and comparing disease control methods, understanding coevolution, and studying disease epidemiology and pathogen ecology. Phytopathometry underpins all activities in plant pathology and extends into related disciplines, such as agronomy, horticulture, and plant breeding. Considering this central role, phytopathometry warrants status as a formally recognized branch of plant pathology. The glossary defines terms and concepts used in phytopathometry based on disease symptoms or visible pathogen structures and includes those terms commonly used in the visual estimation of disease severity and sensor-based methods of disease measurement. Relevant terms from the intersecting disciplines of measurement science, statistics, psychophysics, robotics, and artificial intelligence are also included. In particular, a new, broader definition is proposed for “disease severity,” and the terms “disease measurement” and “disease estimate” are specifically defined. It is hoped that the glossary contributes to a more unified cross-discipline approach to research in, and application of the tools available to phytopathometry.

Information retrieval systems built with a service-oriented architecture have numerous advantages, and portlets are a key technology to implement services which interact with each other in the presentation layer. This work presents an efficient approach for the communication between the components of an information retrieval system based on multiple portlets in a single user interface. It also presents the architecture and the main methods of the system implemented as a case of use for this approach. It is shown that the proposed solution yields the best inter-portlet communication mechanism in each situation, while possessing the ability to deliver aggregated search and superior user experience.

One of the most challenging problems associated with the development of accurate and reliable application of computer vision and artificial intelligence in agriculture is that, not only are massive amounts of training data usually required, but also, in most cases, the images have to be properly labeled before models can be trained. Such a labeling process tends to be time consuming, tiresome, and expensive, often making the creation of large labeled datasets impractical. This problem is largely associated with the many steps involved in the labeling process, requiring the human expert rater to perform different cognitive and motor tasks in order to correctly label each image, thus diverting brain resources that should be focused on pattern recognition itself. One possible way to tackle this challenge is by exploring the phenomena in which highly trained experts can almost reflexively recognize and accurately classify objects of interest in a fraction of a second. As techniques for recording and decoding brain activity have evolved, it has become possible to directly tap into this ability and to accurately assess the expert’s level of confidence and attention during the process. As a result, the labeling time can be reduced dramatically while effectively incorporating the expert’s knowledge into artificial intelligence models. This study investigates how the use of electroencephalograms from plant pathology experts can improve the accuracy and robustness of image-based artificial intelligence models dedicated to plant disease recognition. Experiments have demonstrated the viability of the approach, with accuracies improving from 96% with the baseline model to 99% using brain generated labels and active learning approach.

Colour-thresholding digital imaging methods are generally accurate for measuring the percentage of foliar area affected by disease or pests (severity), but they perform poorly when scene illumination and background are not uniform. In this study, six convolutional neural network (CNN) architectures were trained for semantic segmentation in images of individual leaves exhibiting necrotic lesions and/or yellowing, caused by the insect pest coffee leaf miner (CLM), and two fungal diseases: soybean rust (SBR) and wheat tan spot (WTS). All images were manually annotated for three classes: leaf background (B), healthy leaf (H) and injured leaf (I). Precision, recall, and Intersection over Union (IoU) metrics in the test image set were the highest for B, followed by H and I classes, regardless of the architecture. When the pixel-level predictions were used to calculate percent severity, Feature Pyramid Network (FPN), Unet and DeepLabv3+ (Xception) performed the best among the architectures: concordance coefficients were greater than 0.95, 0.96 and 0.98 for CLM, SBR and WTS datasets, respectively, when confronting predictions with the annotated severity. The other three architectures tended to misclassify healthy pixels as injured, leading to overestimation of severity. Results highlight the value of a CNN-based automatic segmentation method to determine the severity on images of foliar diseases obtained under challenging conditions of brightness and background. The accuracy levels of the severity estimated by the FPN, Unet and DeepLabv3 + (Xception) were similar to those obtained by a standard commercial software, which requires adjustment of segmentation parameters and removal of the complex background of the images, tasks that slow down the process. Made available in DSpace on 2021-09-14T13:01:47Z (GMT). No. of bitstreams: 1 AP-Predictive-models-Forests-2021.pdf: 3767412 bytes, checksum: aeaec795a200faae684d5a3ea753e45e (MD5) Previous issue date: 2021

Pest management is among the most important activities in a farm. Monitoring all different species visually may not be effective, especially in large properties. Accordingly, considerable research effort has been spent towards the development of effective ways to remotely monitor potential infestations. A growing number of solutions combine proximal digital images with machine learning techniques, but since species and conditions associated to each study vary considerably, it is difficult to draw a realistic picture of the actual state of the art on the subject. In this context, the objectives of this article are (1) to briefly describe some of the most relevant investigations on the subject of automatic pest detection using proximal digital images and machine learning, (2) to provide a unified overview of the research carried out so far, with special emphasis to research gaps that still linger, (3) to propose some possible targets for future research.

The severity of plant diseases, traditionally the proportion of the plant tissue exhibiting symptoms, is a key quantitative variable to know for many diseases and is prone to error. Good quality disease severity data should be accurate (close to the true value). Earliest quantification of disease severity was by visual estimates. Sensor-based image analysis including visible spectrum and hyperspectral and multispectral sensors are established technologies that promise to substitute, or complement visual ratings. Indeed, these technologies have measured disease severity accurately under controlled conditions but are yet to demonstrate their full potential for accurate measurement under field conditions. Sensor technology is advancing rapidly, and artificial intelligence may help overcome issues for automating severity measurement under hyper-variable field conditions. The adoption of appropriate scales, training, instruction and aids (standard area diagrams) has contributed to improved accuracy of visual estimates. The apogee of accuracy for visual estimation is likely being approached, and any remaining increases in accuracy are likely to be small. Due to automation and rapidity, sensor-based measurement offers potential advantages compared with visual estimates, but the latter will remain important for years to come. Mobile, automated sensor-based systems will become increasingly common in controlled conditions and, eventually, in the field for measuring plant disease severity for the purpose of research and decision making. Made available in DSpace on 2020-05-12T04:39:44Z (GMT). No. of bitstreams: 1 AP-Phytopathology-Research-2020.pdf: 3687278 bytes, checksum: 1cc4d05c8d98608d3ce2e7d1d2e1b252 (MD5) Previous issue date: 2020 Article 9.

The management of livestock in extensive production systems may be challenging, especially in large areas. Using Unmanned Aerial Vehicles (UAVs) to collect images from the area of interest is quickly becoming a viable alternative, but suitable algorithms for extraction of relevant information from the images are still rare. This article proposes a method for counting cattle which combines a deep learning model for rough animal location, color space manipulation to increase contrast between animals and background, mathematical morphology to isolate the animals and infer the number of individuals in clustered groups, and image matching to take into account image overlap. Using Nelore and Canchim breeds as a case study, the proposed approach yields accuracies over 90% under a wide variety of conditions and backgrounds. Made available in DSpace on 2020-04-16T01:02:40Z (GMT). No. of bitstreams: 1 APCoutingcattle.pdf: 12514186 bytes, checksum: 8780e4b1493c9863882e5fdd2dbebccc (MD5) Previous issue date: 2020 Article number: 2126.

Deep learning architectures like Convolutional Neural Networks (CNNs) are quickly becoming the standard for detecting and counting objects in digital images. However, most of the experiments found in the literature train and test the neural networks using data from a single image source, making it difficult to infer how the trained models would perform under a more diverse context. The objective of this study was to assess the robustness of models trained using data from a varying number of sources. Nine different devices were used to acquire images of yellow sticky traps containing psyllids and a wide variety of other objects, with each model being trained and tested using different data combinations. The results from the experiments were used to draw several conclusions about how the training process should be conducted and how the robustness of the trained models is influenced by data quantity and variety. Made available in DSpace on 2020-10-07T09:14:58Z (GMT). No. of bitstreams: 1 AP-CNN-psyllids-2020.pdf: 4603738 bytes, checksum: 1d3b2dd96f24da3dd75216169a2402f6 (MD5) Previous issue date: 2020

Unmanned aerial vehicles (UAVs) are being increasingly viewed as valuable tools to aid the management of farms. This kind of technology can be particularly useful in the context of extensive cattle farming, as production areas tend to be expansive and animals tend to be more loosely monitored. With the advent of deep learning, and convolutional neural networks (CNNs) in particular, extracting relevant information from aerial images has become more effective. Despite the technological advancements in drone, imaging and machine learning technologies, the application of UAVs for cattle monitoring is far from being thoroughly studied, with many research gaps still remaining. In this context, the objectives of this study were threefold: (1) to determine the highest possible accuracy that could be achieved in the detection of animals of the Canchim breed, which is visually similar to the Nelore breed (Bos taurus indicus), (2) to determine the ideal ground sample distance (GSD) for animal detection, (3) to determine the most accurate CNN architecture for this specific problem. The experiments involved 1853 images containing 8629 samples of animals, and 15 different CNN architectures were tested. A total of 900 models were trained (15 CNN architectures &times, 3 spacial resolutions &times, 2 datasets &times, 10-fold cross validation), allowing for a deep analysis of the several aspects that impact the detection of cattle using aerial images captured using UAVs. Results revealed that many CNN architectures are robust enough to reliably detect animals in aerial images even under far from ideal conditions, indicating the viability of using UAVs for cattle monitoring.

Convolutional Neural Networks (CNNs) usually require large datasets to be properly trained. Although techniques such as transfer learning can relax those requirements, gathering sufficient labelled data to cover all the variability associated to the problem at hand is often costly and time consuming. A way to minimise this challenge would be gathering the training data under laboratory conditions, using high quality sensors capable of generating images with superior resolution, sharpness and contrast. The downside of this approach is that the resulting dataset will most likely lack the variety that can be found under more realistic conditions. This work investigates this trade-off between image quality and dataset representativeness, that is, if a CNN trained with images captured by a scanner in laboratory would be able to reliably recognise psyllids in smartphone images captured under more realistic conditions. A total of 1276 images were used in the experiments, half acquired using a flatbed scanner and half acquired using two different brands of smartphones. Experiments were carried out using Squeezenet CNNs and a 10-fold cross-validation strategy. Accuracies ranged from less than 70% using only scanned images, to around 90% when only smartphone images were employed, indicating that more realistic conditions are essential to guarantee the robustness of the trained network. Scanned images were useful when the training set containing realistic images was not enough to cover all the variability found in the experiments, but were otherwise innocuous.

Deep learning is quickly becoming the standard technique for image classification. The main problem facing the automatic identification of plant diseases using this strategy is the lack of image databases capable of representing the wide variety of conditions and symptom characteristics found in practice. Data augmentation techniques decrease the impact of this problem, but those cannot reproduce most of the practical diversity. This paper explores the use of individual lesions and spots for the task, rather than considering the entire leaf. Since each region has its own characteristics, the variability of the data is increased without the need for additional images. This also allows the identification of multiple diseases affecting the same leaf. On the other hand, suitable symptom segmentation still needs to be done manually, preventing full automation. The accuracies obtained using this approach were, in average, 12% higher than those achieved using the original images. Additionally, no crop had accuracies below 75%, even when as many as 10 diseases were considered. Although the database does not cover the entire range of practical possibilities, these results indicate that, as long as enough data is available, deep learning techniques are effective for plant disease detection and recognition.

Deep learning is quickly becoming one of the most important tools for image classification. This technology is now beginning to be applied to the tasks of plant disease classification and recognition. The positive results that are being obtained using this approach hide some issues that are seldom taken into account in the respective experiments. This article presents an investigation into the main factors that affect the design and effectiveness of deep neural nets applied to plant pathology. An in-depth analysis of the subject, in which advantages and shortcomings are highlighted, should lead to more realistic conclusions on the subject. The arguments used throughout the text are built upon both studies found in the literature and experiments carried out using an image database carefully built to reflect and reproduce many of the conditions expected to be found in practice. This database, which contains almost 50,000 images, is being made freely available for academic purposes.

Over the last few years, considerable effort has been spent by Embrapa in the construction of a plant disease database representative enough for the development of effective methods for automatic plant disease detection and recognition. In October of 2016, this database, called PDDB, had 2326 images of 171 diseases and other disorders affecting 21 plant species. PDDB size, although considerable, is not enough to allow the use of powerful techniques such as deep learning. In order to increase its size, each image was subdivided according to certain criteria, increasing the number of images to 46,513. Both the original (PDDB) and subdivided (XDB) databases are now being made freely available for academic research purposes, thus supporting new studies and contributing to speed up the advances in the area. Both collections are expected to grow continuously in order to expand their reach. PDDB and XDB can be accessed in the link https://www.digipathos-rep.cnptia.embrapa.br/.

The use of hyperspectral imaging (HSI) for deoxynivalenol (DON) screening in wheat kernels is investigated. Experiments were carried out using a new algorithm designed to be simple to implement and computationally light, being largely based on the manipulation of a few selected spectral bands. Initial experimental results revealed that direct estimation of DON content using hyperspectral images is currently unfeasible, but they also indicated that an indirect analysis exploring the correlation between Fusarium damage and DON content may be accurate enough to improve the process of DON screening in the production chain. This motivated the adoption of a classification approach, in which an algorithm, instead of estimating a value for the DON content, classifies wheat kernel batches into two or three categories, depending on the application. The developed algorithm achieved accuracies of 72% and 81% for the three- and two-class classification schemes, respectively. The results, although not accurate enough to provide conclusive screening, indicated that the algorithm could be used for initial screening to detect wheat batches that warrant further analysis regarding their DON content.