Abstract:

Aims: Under the influence of high intensity human activities, e.g. urban expansion, offshore pollutant discharge and marine resources over-exploitation, various ecological problems have been caused in recent years, especially plant species decrease, productivity decline and ecological function degradation in coastal wetland. In this context, a quick and accurate understanding of types and distribution of plant species is of great significance to coastal wetland biodiversity conservation and ecological sustainable development. Due to cloud cover in coastal zones, it is difficult to obtain effective data by using optical satellite remote-sensed images. However, unmanned aerial vehicle (UAV) remote sensing technology can overcome weather constraints and provide intelligent and flexible data acquisition for a feasible technique for plant species monitoring in coastal zones. Compared with pixel-based classification methods, object- oriented classification method can effectively avoid the salt-pepper phenomenon with better classification performance. However, only the low-level features are applied to the object-oriented classification method. Because of this, it is difficult to improve the classification accuracy for complex regions with many plant species. High-level classification features are used by a deep learning method to identify land cover types to achieve higher classification accuracy. In this study, combing UAV remote sensing technology and object-oriented deep learning method, plant species information of coastal wetlands was identified and classified.

Methods: A representative coastal wetland area, located in Minjiang River estuary of South China, was chosen as the research site. High-resolution visible-light images of Minjiang River estuary wetland were obtained by UAV, and field sampling sites were collected by GPS. By the correction, splicing, and clipping of the UAV images, the digital orthophoto map and digital surface model were obtained. Then, on the basis of object-oriented multiresolution segmentation, the optimal segmentation scale of UAV images was determined by the scale parameter estimation model. The optimal feature combination was selected from spectral, texture, shape and height features based on separability analysis. The U-net deep learning method was used to extract the plant species information of coastal wetlands, and its classification accuracy was compared with four machine learning methods: K-nearest neighbor (KNN), decision tree (DT), random forest (RF) and Bayes.

Results: Combing object-oriented method and the U-net classification method, the classification results had the better integrity with less mixing and misclassification than other classification. The object-oriented U-net method could effectively avoid the salt-pepper phenomenon with overall accuracy (OA) 95.67% and the Kappa coefficient 0.91. The OA and Kappa coefficient of each classification were in descending order: U-net > Bayes > RF > DT > KNN. There were significant differences between the producer accuracy and user accuracy of a single plant species in different classification methods. Kandelia candel, Phragmites australis and Ipomoea pescaprae had higher identification accuracy, while Scirpus mariqueter and Cyperus malaccensis had lower identification accuracy.

Conclusion: The object-oriented U-net deep learning method has a favorable classification performance, and its accuracy is significantly higher than other methods in our study. The selection of optimal feature combination is the key to improving the extraction efficiency of coastal wetland plant species information. Our study could provide references for fine classification of coastal wetland plant species, as well as monitoring and biodiversity conservation management of coastal wetlands.

Key words: plant species classification, coastal wetland, object-oriented, U-net deep learning, unmanned aerial vehicle remote sensing, optimal segmentation scale, machine learning