Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Heathlands in Flanders play a crucial role in providing ecosystem services like carbon capture and biodiversity conservation. However, these landscapes have been deteriorating in recent years due to due to altered management practices, land use change and increased nitrogen deposition. This has led to the spread of the invasive grass species, M. caerulea, in these ecosystems. These invasive grasses pose new fire risk in heathlands, further threatening the conservation of these area. Currently, fire risk assessments in Belgium do not consider significant indicators like M. caerulea in heathlands, leading to less accurate risk assessments for heathlands. This study aimed to assess the use of remote sensing techniques to detect M. caerulea in heathlands located in the Campine region of Flanders. Reference data was collected through field surveys in August 2022 and additional data was obtained from the Institute for Nature and Forestry (INBO). This reference data provided information about the percentage coverage of M. caerulea in vegetation plots. The Sentinel-2 network is then used to capture spectral reflectance associated with these plots and calculated vegetation indices such as Normalised Difference Vegetation Index (NDVI) and Normalised Difference Water Index (NDWI). NDVI represents vegetation density, while NDWI indicates the presence of water bodies and vegetation moisture. Using the Sentinel-2 time series data of the NDVI, NDWI, and a combined NDVI and NDWI datasets, Random Forest (RF) classification models was trained to classify M. caerulea coverage in the Campine region. The models' performance was evaluated using the overall accuracy, Cohen Kappa score, recall, precision, and F1-score evaluation metrics. Confusion matrices were used to visualize the distribution of M. caerulea classes in the classifier output. The results showed that the binary classification of M. caerulea coverage (categorized as <5% or ≥35% coverage) outperformed multiclass classification in terms of accuracy and Cohen Kappa scoring. The best performing model used a combined NDVI and NDWI dataset, achieving a 93% accuracy score and a 0.841 Cohen Kappa score. The study also revealed that models trained with data from multiple years performed worse due to inter-year variations in M. caerulea phenology during the summer months. In conclusion, the study demonstrated that Sentinel-2 time series data can accurately detect M. caerulea, but only when classified in a binary format. This binary classification indicates the presence or absence of M. caerulea in a plot but does not provide a comprehensive representation of its extent. This limitation may hinder the assessment of fire risk in heathland areas as it does not fully describe the presence and distribution of M. caerulea.