From Code to Field: Evaluating the Robustness of Convolutional Neural Networks for Disease Diagnosis in Mango Leaves

From Code to Field: Evaluating the Robustness of Convolutional Neural Networks for Disease Diagnosis in Mango Leaves Gabriel Vitorino de Andrade1[0009−0005−6130−8193], Saulo Roberto dos Santos1[0009−0007−8520−8124], Itallo Patrick Castro Alves da Silva1[0009−0008−8543−7776], Emanuel Adler Medeiros Pereira2[0000−0002−6694−5336], and Erick de Andrade Barboza1[0000−0002−0558−9120] 1 Instituto de Computação, Universidade Federal de Alagoas, Maceió, AL, 57072-970, Brazil {gva,srs,ipcas,erick}@ic.ufal.br 2 Centro de Tecnologia, Universidade Federal do Rio Grande do Norte, Natal, RN, 59078-900, Brazil emanuel.pereira.111@ufrn.edu.br Abstract. The validation and verification of artificial intelligence (AI) models through robustness assessment are essential to guarantee the reli- able performance of intelligent systems facing real-world challenges, such as image corruptions including noise, blurring, and weather variations. Despite the global importance of mango (Mangifera indica L.), there is a lack of studies on the robustness of models for the diagnosis of disease in its leaves. This paper proposes a methodology to evaluate convolu- tional neural networks (CNNs) under adverse conditions. We adapted the MangoLeafDB dataset, generating MangoLeafDB-C with 19 types of artificial corruptions at five severity levels. We conducted a benchmark comparing five architectures: ResNet-50, ResNet-101, VGG-16, Xcep- tion, and LCNN (the latter being a lightweight architecture designed specifically for mango leaf diagnosis). The metrics include the F1 score, the corruption error (CE) and the relative mean corruption error (rela- tive mCE). The results show that LCNN outperformed complex models in corruptions that can be present in real-world scenarios such as De- focus Blur, Motion Blur, while also achieving the lowest mCE. Modern architectures (e.g., ResNet-101) exhibited significant performance degra- dation in corrupted scenarios, despite their high accuracy under ideal conditions. These findings suggest that lightweight and specialized mod- els may be more suitable for real-world applications in edge devices, where robustness and efficiency are critical. The study highlights the need to incorporate robustness assessments in the development of intel- ligent systems for agriculture, particularly in regions with technological limitations. Keywords: System Validation · Robustness Assessment · Agricultural AI Systems · Convolutional Neural Networks · Edge Computing · Image Corruption Benchmarks arXiv:2512.13641v1 [cs.LG] 15 Dec 2025 2 G. V. de Andrade et al. 1 Introduction Deep neural networks and machine learning techniques have been widely used in various computer vision tasks, such as object classification. However, unlike humans who can deal with different changes in image structures and styles such as snow, blur and pixelation, computer vision models cannot differentiate in the same way [7]. As a result, the performance of neural networks declines when the images used as input for the model are affected by natural distortions. This highlights the need for system validation and verification to ensure that models perform as expected under different conditions. In production settings, where models will inevitably encounter distorted inputs [14], ensuring thorough system validation and verification processes is crucial. For example, autonomous vehicles must be able to cope with extremely variable external conditions, such as fog, frost, snow, sandstorms, or falling leaves. It is impossible to predict all potential conditions that can occur in nature [9]. Because of this, achieving the kind of robustness that humans possess is an important goal for computer vision and machine learning, as well as creating models that can be deployed in safety-critical applications [7]. Therefore, ro- bust system verification and validation become essential in ensuring that these systems perform reliably. The robustness of models against different types of perturbation has been a much-studied topic in the machine learning community [3]. Natural corruptions, which are an important type of disturbance [3], are com- mon in real scenarios and can reduce the accuracy of models [7], so their study, in conjunction with the validation and verification processes of the system, has been widely carried out [7, 3, 6]. In parallel, modern technologies, including machine learning and computer vision, have been increasingly applied to agriculture to enhance productivity and sustainability [13]. These techniques have introduced innovative trends in monitoring and forecasting [10], which contribute directly to agricultural im- provements [13]. Machine learning models have shown great potential to detect diseases in crop leaves [8], a critical task given that pests and diseases affect an estimated 40% of food crops globally [2]. Among economically important crops, mango (Mangifera indica L.) ranks as the fifth most cultivated fruit world- wide [5], which thrives p

…(Full text truncated)…

This content is AI-processed based on ArXiv data.