Crop-saving with AI: latest trends in deep learning techniques for plant pathology

Abstract

Plant diseases pose a major threat to agricultural production and the food supply chain, as they expose plants to potentially disruptive pathogens that can affect the lives of those who are associated with it. Deep learning has been applied in a range of fields such as object detection, autonomous vehicles, fraud detection etc. Several researchers have tried to implement deep learning techniques in precision agriculture. However, there are pros and cons to the approaches they have opted for disease detection and identification. In this survey, we have made an attempt to capture the significant advancements in machine-learning based disease detection. We have discussed prevalent datasets and techniques that have been employed as well as highlighted emerging approaches being used for plant disease detection. By exploring these advancements, we aim to present a comprehensive overview of the prominent approaches in precision agriculture, along with their associated challenges and potential improvements. This paper delves into the challenges associated with the implementation and briefly discusses the future trends. Overall, this paper presents a bird's eye view of plant disease datasets, deep learning techniques, their accuracies and the challenges associated with them. Our insights will serve as a valuable resource for researchers and practitioners in the field. We hope that this survey will inform and inspire future research efforts, ultimately leading to improved precision agriculture practices and enhanced crop health management.

Keywords: computer vision; deep learning; disease detection; machine learning; plant disease; vision transformers.

Figure 1

A co-word visualization illustrating the research landscape and interplay between computer vision, deep learning technologies and agricultural challenges through an analysis of keywords in research papers.

Figure 2

The structure and organization of the paper at a glance.

Figure 3

Sample images showcasing instances of Apple Scab disease within the Plant Village dataset.

Figure 4

Visual representation of class distribution in the Plant Village dataset, revealing skewed proportions where certain classes dominate a significant portion while others occupy a smaller segment.

Figure 5

Sample images showcasing in-the-wild and lab-controlled instances of soybean within the PlantDoc dataset.

Figure 6

Sample images showcasing instances of Diplodia disease within the Digipathos dataset.

Figure 7

Sample images showcasing instances of Brown Spot disease within the rice disease dataset.

Figure 8

Hand-crafted features framework: image acquisition captures the input. Preprocessing enhances image quality. Segmentation identifies regions of interest. Feature extraction algorithms extract descriptive hand-crafted features. Classification utilizes these features for labeling and predictions.

Figure 9

Image preprocessing: original image, greyscale conversion and soft-edged representation (Left to right).

Figure 10

Visual depiction of image segmentation techniques: Otsu thresholding, background extraction, and foreground/object extraction, showcasing the distinct results achieved through each method.

Figure 11

An overview of R-CNN architecture: RoI warping extracts regions of interest (RoIs) from the input image. A convolutional neural network (CNN) processes these RoIs to extract features. The bounding box regressor refines object locations and sizes. Classification assigns labels to the objects based on the extracted features.

Figure 12

An overview of faster R-CNN object detection pipeline. The RPN generates region proposals, while ROI Projection maps these proposals to feature maps. Finally, ROI Pooling extracts fixed-length feature vectors for classification and bounding box regression for object detection.

Figure 13

Comparison of ResNet residual blocks: (A) Residual block without 1x1 convolution, and (B) Residual block with 1x1 convolution. The addition of the 1x1 convolution in (B) enhances the representation power and allows the network to learn more complex features, leading to improved performance in deep learning tasks.

Figure 14

An illustration of the functionality of a straightforward Fully Convolutional Network (FCN) for precise image segmentation.

Figure 15

An overview of object detection based on DETR.