Rare and Endangered Plant Leaf Identification Method Based on Transfer Learning and Knowledge Distillation

Introduction

Plants are an important part of ecological systems and biodiversity. Protecting plant diversity can benefit the sustainable development of human society (Narzullaev, 2021). Due to the immoderate utilization of wild plant resources, many wild plants are being deracinated rapidly, which severely hurts the balance of the ecological environment on Earth (Fan et al., 2018; Xu et al., 2017). In the meantime, damage to the ecological system is also reducing living space for animals, further accelerating species extinction (Roberts, Hassan, Elamer, & Nandy, 2021). The extinction of each plant species would cause the deracination of 10–30 neighboring species, resulting in a serious impact on the ecological system (Ceballos, Ehrlich, & Raven, 2020). As the number of endangered plants is declining for different reasons, any species extinction could cause damage to the entire ecological system, and the lives and development of humans are threatened. It is necessary to take protective actions as soon as possible to avoid the risk of these extinctions (Li et al., 2018).

Effective protection for rare and endangered plants includes in situ conservation, namely, to build nature reserves in the primitive environment and reduce artificial interference and damage to realize the goal of protecting these species (Mounce, Smith, & Brockington, 2017). Due to the wide required scope of knowledge, relevant personnel are facing difficulties in identifying endangered plants. This escalates the degree of difficulty of endangered plant protection. Significant efforts and follow-up are urgently needed (Liu, 2021). It has become important to mitigate germplasm resource loss caused by difficulties for personnel in endangered plant identification by adopting automatic identification based on computer, artificial intelligence, and other similar technologies.

Blooms, leaves, and fruit could all be used as evidence for endangered plant identification. Among these parts, leaves have become the major method for plant identification, as they are stable in structure, easy to be collected, and have long life cycles. However, artificial identification of leaves requires a lot of work and is inefficient. It is a subjective method and requires a wider scope of knowledge than other identification practices. So far, there has already been some research on automatic leaf identification methods. These methods mainly identify leaves by their superficial characteristics, such as color, shape, and texture (Saleem, Akhtar, Ahmed, & Qureshi, 2019; Singh, Nagahma, Yadav, & Yadav, 2018; Yang, 2021). Most plant leaves are green, which means there is no obvious difference between plants. In contrast, shapes do have obvious differences, even on the same plant. Textures are too difficult to describe and to use to yield conclusions. Therefore, these adopted superficial characteristics lack universality. Models calculated based on such characteristics are also incapable of being used for generalization. It has become common for recent research to automatically extract deep features of leaves through deep learning (Lee, Chan, Mayo, & Remagnino, 2017; Lin, Ding, Tu, Huang, & Liu, 2019). Due to the small sample size of plants, especially endangered plants, in real applications, the direct adoption of the deep learning model may lead to overfitting. Transfer learning is thus used in the plant leaf identification model (Gogul & Kumar, 2017; Kaya et al., 2019). However, the large number of parameters created for these models after transfer learning requires more computing resources. In outdoor environments and other scenes without network and server support, it would be difficult to apply them to smart mobile terminals and other technologies with low computing power. Moreover, no research specially focuses on rare and endangered plant leaf identification.

In response to these issues, this paper proposes an endangered plant leaf identification method based on transfer learning and knowledge distillation, which guarantees high accuracy and requires fewer convolutional neural network (CNN) parameters. The lightweight endangered plant leaf identification model is proposed for application to more settings. The major contributions of this paper are as follows: