Vegetation Cover Classification of Coastal Sand Dune Ecosystems Using Ultra-High-Resolution UAS Imagery and Machine Learning Techniques
Elena Belcore, Melissa Latella, Marco Piras, Carlo Camporeale
The increasing use of Uncrewed Aerial Systems (UAS) has opened up new opportunities for ultra-high-resolution (UHR) land cover (LC) classification using optical data with Ground Sampling Distance (GSD) below 10 cm. Coastal sand dune ecosystems are difficult to map due to the variability of plant species, making high-resolution vegetation mapping of these areas crucial for analysing vegetation dynamics, spatial patterns and predicting species diversity. The extreme similarity of vegetation spectral responses to multispectral sensors, the small size of the coastal dune plants (mostly herbaceous), and the large amount of data generated are the main challenges in achieving ultra-high-resolution LC maps of vegetation mapping.
This work focuses on developing a VHR vegetation cover classification model for three areas of the San Rossore National Park in Italy using data collected by UAS (DJI Phantom 4 multispectral) with a multispectral optical sensor (RGB, Redge, NIR). The machine learning model is trained on two phenological-relevant epochs (September 2021 and May 2022) using a sampling scheme that combines UAS flight acquisition and field vegetation survey data collected at high precision positioning (dual frequency GNSS). A total of 757 herbaceous and shrub species were sampled.
The VHR classification of 12 species and 2 service classes (Debris and Sand) is multitemporal supervised object-oriented (OBIA), characterised by spectral features, spectral indices, elevation, and texture. Three areas of about 5 hectares each were analysed, one used solely for transferability tests.
The calibrated multispectral orthomosaics and the Crown Height Model (CHM) were generated with Structure from Motion-based processing. Textural features based on Haralick co-occurrence matrix and spectral indices were computed, resulting in a final dataset of 31 features.
The semantic segmentation was performed using eCognition Developer (Trimble), based on the Normalised Difference Vegetation Index (NDVI), RGB and CHM of May 2022 dataset, resulting in 383’200 elements over the three study areas. Imbalanced datasets, such as the one of this work, may lead to inaccurate classification, so the borderline synthetic minority oversampling technique (SMOTE) was used for oversampling the training dataset.
The random forest algorithm was used to classify tree species, and feature selection based on GINI impurity was conducted to reduce the dimensionality of the input features (reduced to 19 based on the statistical distribution of impurity).
To verify the accuracy of the model, a primary accuracy measure based on the error matrix was calculated, and the model was cross-validated using a 100-fold stratified cross-validation. The overall accuracy (OA) was found to be 0.77, with a standard deviation of 0.14. After feature selection, the OA slightly decreased to 0.76, but the processing time was improved, and the standard deviation was reduced to 0.13. The model was then applied to an unseen dataset of the transferability-test area, and the OA decreased to 0.62.
In conclusion, using UAS and multispectral ad multi-temporal optical data provides a valuable tool for ultra-high-resolution LC mapping of vegetation in challenging environments such as coastal sand dunes. The developed vegetation cover classification model based on machine learning algorithms accurately classifies vegetation species and its performances are in line with the literature. Further research is needed to improve the model's accuracy when applied to different datasets and to extend the model to map other vegetation-dominated dune environments.