Automated classification of saffron and broadleaf weeds of Flixweed and Hoary Cress using deep learning and color images

Abstract

Today, the use of precision agriculture strategies that lead to effective weed management and minimize damage to crops is of great importance. In this research, color images and deep learning were used to distinguish Saffron from two common broadleaf weeds, namely Hoary Cress and Flixweed. A total of 291 photos of the three classes were taken under completely natural field conditions. The image data were resized to 150 × 150 pixels, normalized to [0,1] range, and then augmented. A total of 24 deep learning models with varying architectures and parameters were evaluated to identify the most optimal structure. The proposed models were established on the Xception network and leveraged initial weights from ImageNet via the transfer learning approach. These models underwent fine-tuning and modification by introducing additional layers to the base model. Ultimately, two models were proposed based on their performance in accuracy and loss on test data. One model demonstrated impeccable performance, attaining 100% accuracy and a loss of 0.41, despite potential network instability. It successfully differentiated all classes with precision, recall, and F1-score all at 100%. The second model proposed exhibited a more stable performance with 95% accuracy and a loss of 3.78. It achieved precision ranging from 91% to 100%, recall from 88% to 100%, and F1-score from 91% to 96% across the three classes, showing promising results. The results of this research can be considered as a basis for the development of weed removal robots for the purpose of automatic weed control using precision agriculture strategy.