Abstract
Many areas across Europe are mapped and monitored using a large range of different data types, sources and classification schemes leading to gaps in the knowledge required to fulfill the European Council’s Habitats Directive (1992). The Earth Observation Data for Habitat Monitoring (EODHaM) system, developed during the EU FP7 BioSOS project, introduces a systematic, hierarchical approach that is applicable to all sites and available as a standard, providing classifications of high value for conservation and biodiversity purposes (Lucas et al. Int J Appl Earth Observ Geoinf 37:17–28, 2015). The system is built on the Land Cover Classification System (LCCS) developed by the FAO for use in the field. The aim of this project is to generate accurate maps of the location, extent and condition of coastal Annex I habitats at Kenfig Burrows Special Area of Conservation (SAC), using VHR Worldview-2 data.
Indices, such as Normalized Difference Vegetation Index (NDVI) allow straightforward visual threshold determination in the rule base, classifying LCCS Level 3 with accuracies of 90% and above. Beyond Level 3, in situ data is key for training and validating EO data to determine if (a) lifeforms/habitats are separable with the available EO data, and (b) suitable thresholds can be determined for classification. Numerous indices can be calculated, and using the GPS point training data, a separability analysis based on Analysis of Variance (ANOVA) allows those with the highest separation scores to be chosen as layers for classification. By plotting the training data sets into boxplots, suitable thresholds are determined. The appropriateness of LCCS here varies with specific sites; for example, slack habitat in sand dune ecosystems can be accurately mapped from contextual information derived from slope (calculated using VHR LiDAR data) and can therefore be translated to habitat from LCCS Level 3. Classifications are therefore translated from land cover to habitat after LCCS Level 3 instead of following the hierarchy to Level 4 and beyond.
Once the broad habitat baseline is mapped, thresholds become restricting as they set clear straight lines in the feature space when classifying, therefore machine learning techniques such as random forest and/or support vector machines are more suitable for determining whether dominant species within broad habitat classes can be separated and classified accurately. By classifying dominant species, condition of habitats can be inferred. With accuracies of classifying some habitats higher than others when implementing EO data into a monitoring system, field surveying can never be ruled out to attain the knowledge required for the habitats directive. However, surveying can be applied specifically to those habitats that EO data cannot sufficiently classify
Indices, such as Normalized Difference Vegetation Index (NDVI) allow straightforward visual threshold determination in the rule base, classifying LCCS Level 3 with accuracies of 90% and above. Beyond Level 3, in situ data is key for training and validating EO data to determine if (a) lifeforms/habitats are separable with the available EO data, and (b) suitable thresholds can be determined for classification. Numerous indices can be calculated, and using the GPS point training data, a separability analysis based on Analysis of Variance (ANOVA) allows those with the highest separation scores to be chosen as layers for classification. By plotting the training data sets into boxplots, suitable thresholds are determined. The appropriateness of LCCS here varies with specific sites; for example, slack habitat in sand dune ecosystems can be accurately mapped from contextual information derived from slope (calculated using VHR LiDAR data) and can therefore be translated to habitat from LCCS Level 3. Classifications are therefore translated from land cover to habitat after LCCS Level 3 instead of following the hierarchy to Level 4 and beyond.
Once the broad habitat baseline is mapped, thresholds become restricting as they set clear straight lines in the feature space when classifying, therefore machine learning techniques such as random forest and/or support vector machines are more suitable for determining whether dominant species within broad habitat classes can be separated and classified accurately. By classifying dominant species, condition of habitats can be inferred. With accuracies of classifying some habitats higher than others when implementing EO data into a monitoring system, field surveying can never be ruled out to attain the knowledge required for the habitats directive. However, surveying can be applied specifically to those habitats that EO data cannot sufficiently classify
Original language | English |
---|---|
Title of host publication | The Roles of Remote Sensing in Nature Conservation |
Editors | Ricardo Díaz-Delgado, Richard Lucas, Clive Hurford |
Publisher | Springer Nature |
Pages | 91-120 |
ISBN (Electronic) | 978-3-319-64332-8 |
ISBN (Print) | 978-3-319-64330-4, 3319643304 |
Publication status | Published - 10 Dec 2017 |
Keywords
- annex I
- habitat mapping
- Land Cover Classification System
- EODHaM system
- machine-learning