Plant diseases recognition on images using convolutional neural networks: A systematic review

Potato crop stress identification in aerial images using deep learning‐based object detection

“…FN denotes the false negatives , that is, the set of pixels in an image that are incorrectly predicted by the model to belong to the negative class. Precision calculates the percentage of the pixels in an image that are correctly predicted by the model as belonging to the positive class out of all the predicted pixels by the model, that is, Recall (also known as sensitivity) expresses the percentage of the pixels in an image that are correctly predicted by the model as belonging to the positive class out of all relevant pixels that belong to the positive class according to the ground truth, that is, Although in most related works for object detection (Abade et al., 2020; Nagaraju & Chawla, 2020) the employed metrics–such as mean Average Precision (mAP)–are calculated based on the number of correctly predicted bounding boxes for the objects of interest, here we use metrics that are calculated based on correctly predicted pixels in crop images for healthy and stressed plants. The reason for such choice of metrics stems from the difficulty in distinguishing potato plants as single objects when the canopy is closed.…”

“…Frequently applied image analysis methods encompass support vector machines (Camargo & Smith, 2009; Rumpf et al., 2010), k ‐nearest neighbors, random forests (Lottes et al., 2017), and multivariate Gaussian classifiers (Hamuda et al., 2016). However, conventional machine learning approaches rely on manually fine‐tuning a set of parameters for a given collection of images, which may lead to decreased performance on images taken under different environmental conditions (e.g., illumination, weather), growing stages, or soil types (Abade et al., 2020; Milioto et al., 2017).…”

“…A recent line of research emphasized the advantages of deep learning (DL)‐based methods for image analysis in PA (Abade et al., 2020; Ghosal et al., 2018; Nagaraju & Chawla, 2020; Ramcharan et al., 2017; Taghavi Namin et al., 2018; Yamamoto et al., 2017). Contrary to the conventional methods that typically employ steps of segmentation, feature extraction, and classification, DL models comprise multi‐layer architectures that integrate all processing steps into one single framework for automated feature extraction and mapping the inputs to an output function (Goodfellow et al., 2016; LeCun et al., 2015; Zhang et al., 2018).…”

“…Still, preparing ground truth labels for image segmentation requires demarcating all objects in images, which is substantially more time‐consuming in comparison to defining the location of a set of bounding boxes in object detection. Subsequently, object detection methods have been the preferred choice in many related tasks (Abade et al., 2020; HamidiSepehr et al., 2019) where the approximate locations of the bounding boxes for the objects of interest provide a satisfactory level of accuracy.…”

Inhibitory Activity of Shrimp Waste Extracts on Fungal and Oomycete Plant Pathogens

“…In this respect, the losses of products along the post-harvest chains (i.e., warehousing, transport and final distribution) determine strong impactful consequences, especially in agriculture-based-economy countries [1-3]. To minimize production losses and maintain crop sustainability, several strategies based on the application of different means, such as physical, chemical and biological, have been adopted over time [4,5]. Currently, one of the most consolidated and effective means for controlling fungal diseases is represented by chemical synthetic fungicides [4,6].…”