Attention-enhanced VGG-16 architecture for precision weed detection in wheat fields

Abstract

This study delineates five advanced convolutional neural network designs that employ deep learning for the classification of wheat weeds. The dataset consists of greyscale photos taken in agricultural fields, enhanced with RGB lighting and cropped to 256×256×3 pixels. Normalization, batch-wise augmentation, contrast stretching, or histogram equalization were all part of the preprocessing that improved picture quality and model learning efficiency. These enhancements developed feature extraction by increasing picture contrast and homogeneity. A VGG16-based model with further Conv2D layers and spatial attention is one of the five models recommended. An additional design, inspired by ResNet50 that applies residual blocks and worldwide average pooling. A hybrid RNN integrated ICNA-CNN or LSTM for spatial-temporal content learning is another. Lastly, InceptionResNetV2 is enhanced with CNN layers and a query-key attention mechanism. We used the Adam optimizer and categorical cross-entropy to estimate the loss throughout the training for the round. At the end, all the models had ReLU activation, batch normalization, MaxPooling2D, and a dropout charge of 0.5 to keep them from overfitting. Upon evaluating their performance based on accuracy, recall, reliability, and training loss, VGG16 emerged as the standout model, achieving an impressive 99.07% precision and a remarkably low training loss of just 0.0079. The study determined that VGG16 is the optimal choice for precise wheat–weed categorization because to its superior generalizability and accuracy.