From Lab to Field: An Empirical Study on the Generalization of Convolutional Neural Networks towards Crop Disease Detection

Abstract

Due to complex feature abstraction and learning power, CNNs have been the most successful machine learning algorithms for image classification tasks. The objective of this work was to evaluate the potential of convolutional neural networks (CNNs) for extracting underlying complex features and recognize these patterns towards the task of detecting healthy and diseased crop plants. The generalization of these algorithms was assessed on different situations of training and testing scenarios using images from controlled lab conditions and real field environments. Results have shown that when presented with sufficient data variability in training, englobing images with similar conditions faced in testing, the deep learning architectures delivered accurate results of over 90%. In contrast, the same architectures were not able to generalize the accuracy of training towards the detection of new unseen images that were not extracted in the same settings as the ones from the training set, delivering, in this case, a general accuracy of around 50%. The deployment of practical automated support systems for disease detection depends on the provision of robust datasets for training CNNs which contemplate the spectral variability conditions found in numerous crop cultivation environments encountered in diverse field sites across the globe.

Embracing limited and imperfect training datasets: opportunities and challenges in plant disease recognition using deep learning

“…Second, the distribution of the training dataset is gradually approaching that of the test dataset when the training dataset becomes larger, supporting a better test performance. For example, a model trained with images captured in laboratories is not expected to be effective when tested with images captured on farms ( Guth et al., 2023 ; Wu et al., 2023 ).…”

“…A frequent performance drop occurs when a model is trained on a dataset from a particular scenario but is further tested on data from a different scenario. A common scenario, for example, is that the training dataset is collected in one place by one person and the test dataset is collected in another place with different infrastructures and illumination by another person with their individual habit of taking pictures, such as training in the images collected in the laboratory and testing in the real world ( Guth et al., 2023 ; Wu et al., 2023 ).…”

“…Another essential aspect to investigate in the domain shift is the changes that occur in data collected in the laboratory compared with data collected under field conditions. The generalization capabilities of CNNs are investigated to learn the clear patterns from lab conditions that are to be detected again under new and more complex field conditions while avoiding overfitting ( Guth et al., 2023 ). The important insight derived from this is that, in order to create useful tools for disease detection and classification using deep learning for image analysis, it is crucial to develop a final product that can handle a wide range of images from various crop conditions and locations, including inter-class and intra-class variations ( Wu et al., 2023 ).…”