Explainability of Deep Learning-Based Plant Disease Classifiers Through Automated Concept Identification

Jihen Amara, Birgitta König-Ries, Sheeba Samuel · 2024

Read Explainability of Deep Learning-Based Plant Disease Classifiers Through Automated Concept Identification by Jihen Amara, Birgitta König-Ries, Sheeba Samuel (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-07408.txt.

Introduction

Jihen Amara, Birgitta König-Ries, and Sheeba Samuel ask a plant disease model to show its work in a 2024 paper. Deep learning spots crop disease from leaf photos with high accuracy. Farmers and agronomists, though, cannot trust a verdict with no reasons.

An automated concept method mined the model's own image collection for the visual ideas it uses. On a standard network trained on the PlantVillage set, the method surfaced true disease patterns. It also caught the model leaning on background and lighting, biases that can break it in real fields. Each found bias is a repair target. The authors pitch concept audits as routine for farm AI tools.

Catalog record

CREATORJihen Amara, Birgitta König-Ries, Sheeba Samuel
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-07408

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Explainability of Deep Learning-Based Plant Disease Classifiers Through Automated Concept Identification Jihen Amara, Birgitta König-Ries, Sheeba Samuel Abstract While deep learning has significantly advanced automatic plant disease detection through image-based classification, improving model explainability remains crucial for reliable disease detection. In this study, we apply the Automated Concept-based Explanation (ACE) method to plant disease classification using the widely adopted InceptionV3 model and the PlantVillage dataset. ACE automatically identifies the visual concepts found in the image data and provides insights about the critical features influencing the model predictions. This approach reveals both effective disease-related patterns and incidental biases, such as those from background or lighting that can compromise model robustness. Through systematic experiments, ACE helped us to identify relevant features and pinpoint areas for targeted model improvement. Our findings demonstrate the potential of ACE to improve the explainability of plant disease classification based on deep learning, which is essential for producing transparent tools for plant disease management in agriculture. 1 Introduction Agriculture is primordial for human life. It is the foundation for food production and economic stability around the world. With the global population expected to reach 10 billion by 2050, food production must increase by 60% to meet growing demand [ 26 ] . To achieve this, it is required to boost crop yield quality and minimise food loss caused by diseases. Hence, efficient and innovative agricultural practices are more important than ever. Consequently, in recent years, we have noticed the emergence of automated plant disease detection methods that can save time and provide better help and support to farmers and other stakeholders. Central to these innovations is deep learning (DL), a powerful branch of AI that enables the rapid and accurate identification of plant diseases through image analysis. By detecting diseases early and with greater precision, deep learning enhances the effectiveness of these automated systems. This can allow for more timely and targeted interventions, ultimately helping to save crops and reduce losses. Different works have applied deep learning for plant disease classification and demonstrated its efficiency. They often employ convolutional neural networks (CNN) such as inceptionV3 [ 29 , 9 , 20 , 24 , 28 ] , which are suitable for image classification tasks. Even though deep learning methods have advanced image-based plant disease classification, they often lack explainability, which makes users and involved stakeholders question their results. Deep learning models are usually referred to as a black box due to the complexity of their architectures and the lack of insight into how they reach their decisions. This opacity is a huge problem in critical domains like healthcare, finance, or agriculture, where decisions come with high stakes and the threat of significant financial losses. Hence, understanding and explaining the rationale behind the model predictions is essential for farmers and agricultural experts to trust the model’s output especially when it is related to disease management strategies. Also, explainability can help in detecting and preventing unexpected biases, especially when dealing with noisy and unbalanced datasets which is often the case in plant disease classification [ 3 ] . One of the most popular explainability methods is the visual-based approach which helps users understand model decisions by visually highlighting the areas or features in the input data that contribute most to the model’s predictions (see Section 2 ). Even though these visual based explanation methods provide useful visual feedback, they often fail to capture more abstract or conceptual information that influences model decision-making. To address this limitation, we propose the use of concept-based explainability methods. Specifically, Automated Concept-based Explanation (ACE) [ 12 ] stands out as a method that can identify and group high-level concepts directly from the data. Unlike visual based methods, ACE can help by quantifying the importance of high-level human-interpretable concepts such as leaf color, shape, or disease shape and texture in the model’s decision-making process. This helps agriculture experts check whether the model is generating outcomes based on valid biologically relevant features. Also, we argue that ACE could not only be a great tool to explain the model decision but can be a valuable method to explore biases and detect irrelevant concepts in the dataset. In some cases, the model may start learning irrelevant patterns or correlations to make its decision due to biases in data collection (i.e background noise). This can result in poor generalisation and decreased accuracy in real-world applications. Finally, our contributions can be summarised as follows: • We provide new and valuable insights into the DL-based plant disease classification models by adapting the ACE method to automatically identify key visual patterns and concepts used in the model’s decision-making process. This approach allows plant and agricultural experts to better understand the model’s behaviour, raising greater trust in its predictions. • We introduce a new approach that reveals biases in DL model-based plant disease classification and identifies potential issues within the dataset using ACE. This approach helps diagnose if the model’s decisions rely on unwanted correlations instead of meaningful plant disease symptoms. This helps to address and mitigate these issues in future models and datasets toward a more robust and generalizable plant disease classification. • We offer technical insights into the application of ACE in the context of plant disease classification, providing a crucial tool for identifying and addressing unique challenges within this field. To the best of our knowledge, this is the first work to apply ACE in this domain, establishing an important step toward improving the reliability and explainability of DL models for plant disease diagnosis. The remainder of this paper is organised as follows. Section 2 presents related work on explainability and its application to plant disease classification. Section 3 details the model, ACE theory, and dataset used in this study. Section 4 presents our experimental results, including discussions on prediction accuracy and model explanation with ACE. Finally, Section 5 offers concluding remarks and possible future work. 2 Related work Several explainability methods have been proposed, with the most common approaches falling into two categories: visual-based explanations and concept-based explanations. In this section, we will describe each set of methods and provide examples of their use in the context of plant disease classification. Visual-based explanation [ 32 , 27 , 25 ] generates visual cues highlighting the parts of an input image that contribute most to the model’s prediction [ 22 ] . These methods produce heat maps or saliency maps that indicate which areas of an image the deep model considers important when making its predictions. Zhou et al. [ 32 ] proposed Class activation mapping (CAM), a technique that replaces the fully connected layers at the end of a CNN with a global average pooling layer (GAP), applied on the last convolutional feature maps. The CAM is then computed as a weighted linear sum of these feature maps, where the weights are determined by the output class probabilities of the CNN. This resulted in a heatmap that highlighted the regions of the image that were most strongly associated with the predicted class. Selvaraju et al. [ 27 ] proposed Gradient-weighted Class Activation Mapping (Grad-CAM), a generalisation of CAM that can work with any type of CNN to produce local explanations. This is in contrast to CAM, which specifically needs global

Read the complete text