Narrow-leaved lupin seed variety identification with weighted ensemble soft voting technique

Narrow-leaved lupin ( Lupinus angustifolius ) is an important legume in Australia due to its high protein content, applications in human food as well as livestock feed, and its ability to improve soil fertility through nitrogen fixation. Lupin varieties have different adaptation and quality profiles for production and market preference. Therefore, quick and accurate variety identification from seed is important for crop management, breeding, seed quality assessment and marketing. However, many lupin varieties show highly similar visual characteristics, making manual identification difficult and time-consuming. The objective of this study is to develop and evaluate deep learning (DL) techniques for automated lupin variety seed identification using RGB images. Accordingly, a dataset was created comprising 12 distinctive lupin varieties which includes PBA Barlock, Tanjil, PBA Gunyidi, Mandelup, Jenabillup, Quilinock, PBA Jurien, Unicrop, Coromup, Kalya, Rosemont, and Danja. The dataset contains both healthy and Phomopsis diseased lupin seed images. Several deep learning models including EfficientNetB0, InceptionV3, SqueezeNet, ShuffleNet, MobileNetV2, ResNet50, and DenseNet121 have been trained and tested with this dataset. Among all, the weighted ensemble of ResNet50 and DenseNet121 outperformed the other models with an accuracy of 98.85%. Results show that DL models can effectively identify visually similar lupin variety seed despite challenges such as inter-class similarity and intra-class dissimilarity. The proposed ensemble model presents a pioneering approach for automated lupin variety seed identification. This can support farmers, marketers and breeders in improving crop quality through better assurance of seed purity.