Early Fusion of Discrete Wavelet Transform Features and Convolutional Neural Networks for Precise Detection of Apple Plant Diseases

Apple plant diseases significantly reduce agricultural productivity, necessitating the need for precise detection methods. To this end, in this paper, we propose a novel approach that combines Convolutional Neural Networks (CNNs) with Discrete Wavelet Transform (DWT) descriptors using high-low fusion strategyFeatures are used for improvement in disease detection. In particular, DWT decomposes images into spatial-frequency components, which are used to describe texture and edge features which are important to identify illness symptoms. These are concatenated with original image data at the input layer in the CNN, allowing for space and frequency domain information to be used simultaneously. The data was evaluated on a systematic approach of open dataset of the apple plant disease by using the 4 -fold cross validation to make sure that a diversity and robustness of the model. The results show a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$7.4 \%, 6.8 \%$</tex> and 6.2 % increase in accuracy, F1-score and sensitivity when compared to regular CNN models. The approach was also more resilient to challenging conditions such as noise and variations in illumination. We also compared performance of models including AlexNet, VGG-16, EfficientNet, and ResNet50. This confirmed that early fusion of DWT features aided in detecting these time series, as our approach was found outperforming the models by higher margins. Ablation studies confirmed that the effectiveness of the proposed model was mainly attributed to the addition of wavelet features at the input stage. This study opens the door to more effective precision agriculture solutions by highlighting the potential of wavelet transforms in conjunction with deep learning for precise, scalable agricultural disease identification.