• Proposed a segmentation method for extracting plant parts from complex environments with low lighting conditions. • Compared the proposed segmentation model against other existing segmentation methods. • Different deep learning models are trained on segmented and non-segmented images for classifying various seedlings to accurately determine an effective transfer learning model with the best possible recognition rate. Early identification of plant species in uncontrolled conditions can be complex due to the complex background and morphological diversity of a plant's leaves. Agriculture plays an important role in our daily life to produce food, develop economies, international trade, research, and promote global sustainability. In this work, classification of plant species is performed based on a proposed segmentation method using a colour thresholding technique to separate the plant regions in images captured in low lighting and a complex background. The models include DenseNet121, NasNet Mobile, MobileNet, RegNet, and XceptionNet, trained on the proposed segmentation method and raw image samples, and compared for their effectiveness. The dataset, comprising 4,750 images of 12 different types of plants, their seedlings, and weeds, is sourced from Kaggle. The results demonstrate that the model's classification performance using the proposed segmentation method consistently outperformed with the non-segmented images, confirming the need for segmentation to improve model accuracy.
R et al. (Sun,) studied this question.