Research On Mangrove Species Classification Integrating Airspace Remote Sensing And Adaptive Ensemble Learning

Mangroves are one of the most productive wetland ecosystems in the tropical and subtropical coastal land-sea interface,and accurate classification of the tree species is crucial for mangrove management and conservation.This paper selects several mangrove reserves in Beibu Gulf as the research object,combination of multispectral UAV images,optical satellite images(GF-7、ZY302 and Sentinel-2A)and GF-3 SAR images as data sources,and uses Stacking ensemble learning(SEL)and deep learning algorithms to establish mangrove species classification models and explore the feasibility of SEL algorithm for identifying the mangrove species.This paper designs two novel adaptive Stacking ensemble learning algorithms(ASEL)to evaluate the adaptive correlation elimination Stacking(ACE-Stacking)algorithm and adaptive optimal model Stacking(AOM-Stacking)algorithm for mangrove species classification ability and generalization based on the relevance and classification accuracy of the base model.The backscatter coefficients and polarization decomposition parameters of GF-3 with dual and full polarization were extracted to explore the effects of SAR images with different polarizations on the classification of mangrove species,and to demonstrate the adaptability of ASEL algorithm for large-scale mangrove species identification.The results of the study show that:(1)The traditional SEL algorithm has strong performance in mangrove species classification.the SEL algorithm has better classification capability than RFE-Deeplab V3+ and RFE-PSPNet algorithms,and the overall classification accuracy is 1.6%-12.7% higher than the two deep learning algorithms.(2)The traditional SEL algorithm is superior to the single machine learning algorithm for mangrove species recognition.Among the six classification schemes,the overall classification accuracy of the SEL algorithm is 75.5%-92.2%,and the highest overall classification accuracy is 0.8%-4.2% higher than that of the single classification algorithm,and the average accuracy of all three types of mangrove species is above 0.87.Texture features can significantly improve the classification accuracy of Kandelia candel,and DSM features can significantly improve the classification accuracy of Avicennia marina and Aegiceras corniculatum.(3)Both adaptive Stacking ensemble learning algorithms can achieve high accuracy classification of mangrove species with strong classification and generalization ability.The overall classification accuracies ranged from 79.8%-96.2% in the four aerial photography areas.Among them,the ACE-Stacking and AOM-Stacking algorithms have an mean overall classification accuracy of 0.9%-3.3% higher than that of the traditional Stacking algorithm.The average accuracies of both ASEL algorithms for seven mangrove species were above 93%.(4)There were differences in the identification accuracy of mangrove species in different protected areas,and the F1 score values mangrove species in the four areas are 84.8%-98.4%.The Cyperus malaccensis have the highest recognition accuracy and Avicennia marina had the lowest accuracy.The SHAP method explained that the spectral band and vegetation indices were of high importance for the classification accuracy of the model.in addition,the red band in the spectral band had a high contribution to Aegiceras corniculatum and the ratio vegetation index had a high contribution to Cyperus malaccensis had a higher contribution.(5)The classification accuracy of mangrove species increases with the increase of spatial resolution of remote sensing images.All-polarized SAR and VV-VH-polarized SAR images significantly improved the classification accuracy of mangrove species.The average mangrove species accuracy of the combination of polarized SAR image features improved by 2.4-4.9%compared with that of using only spectral data.