Tree species classification NRW

Description

Classification of tree species groups at stand level, based on Sentinel2 - satellite images The classification into the 9 most frequently occurring tree species groups in North Rhine-Westphalia takes into account the following classes: Oak (egg), beech (Bu), coloured leaves (hALH), soft leaves (wALN), poplar, pine, larch, spruce and Douglas fir (see legend). The classification was carried out by a method of artificial intelligence (recurrent neural network), which was trained on the basis of forest facility data from the forest areas of the state-owned forestry company in NRW. The procedure was developed on behalf of the Ministry of Agriculture and Consumer Protection of the State of North Rhine-Westphalia (MLV NRW). In order to identify the tree species on the basis of their phenology and to ensure cloud-free conditions, time series of satellite images from the growing seasons of two years between 03.03.2022 and 11.07.2023 were evaluated. A total of 516 individual scenes were selected and assembled nationwide into a cloud-free multitemporal mosaic (https://www.d-copernicus.de/daten/satelliten/satelliten-details/news/sentinel-2/ ). Due to the spatial resolution of the Sentinel2 satellite images of about 10 m, this map plane is suitable for viewing on a scale level of up to about 1:10,000, not larger! The presentation takes into account only tree species in the main layer and is not comparable to a forest establishment or forestry operation map. Mixtures of tree species are characterized by changing colors even within a stock. The data on the bares are based in particular on the aggregated calamity map in the coniferous forest (status: Sep 2023). Validation User accuracy is the key to the reliability of map information in the remote sensing-based tree species map. The user accuracy describes a representative estimate of the probability that the map information shown in the tree species map (tree species group) at a location coincides with the main layer of a stock there. The user accuracy is different for each designated thematic class (tree species group) and is determined by the proportion of the above-described reference data from the forest facility with the corresponding tree species group in the main layer, in relation to all reference points in the validation sample that have been classified by the neural network into the corresponding tree species group. The accuracies therefore differ between the individual tree species depending on the frequency with which the tree species are represented in the forestry facility. Frequent tree species are comparatively harder to detect for an artificial intelligence lighter, less frequently occurring tree species (e.g. poplars, Douglas fir). The manufacturer's accuracy is used to assess the classification quality based on the proportion of reference data in the validation sample with the corresponding tree species group correctly assigned to this tree species group by the neural network. The F-measurement as a general class-specific quality measure is formed by the harmonic mean (average) of user and manufacturer accuracy. The overall accuracy, regardless of class affiliation, indicates the proportion of reference points in the validation sample that has been correctly classified by the neural network. A common and widely accepted measure of quality is Cohen’s κ coefficient, which describes the correspondence between the class information in the reference data and the classifications of the neural network based on the validation sample. In practical application, the range of values of the coefficient is between 0 and 1, where a value of 1 means a complete agreement of the results, while a value of 0 means a purely random agreement without any discernible added value of a decision of the neural network. User Accuracy (%) Oak Beech hALH wALN Poplar Pine Larch Spruce Douglas fir 93.8 94.7 91.2 89.2 88.8 97.7 95.9 97.1 88.4 Manufacturer Accuracy (%) 91.0 96.0 91.4 92.7 79.0 97.5 90.7 98.0 76.4 F-Measurement (%) 92.3 95.4 91.3 90.9 83.6 97.6 93.2 97.6 82.0 Total Accuracy: 95,0 % Cohen’s κ coefficient: 0.938 Table 1: Accuracy data from statistically independent validation. The specified accuracies were determined from a statistically independent validation based on the forest establishment data from the state forests. For this purpose, the classification result in more than 15,000 randomly selected forest stands nationwide was compared with the uniform tree species in the main stratum. This process was repeated three times on different figurative elements. The percentages given indicate the relative frequency (statistical probability) with which the classification result and the tree species reported in the main stratum coincide. As a prerequisite for independent validation, there is no overlap between training and validation data (statistical independence).

Resources

Name Format Description Link
0 https://www.opengeodata.nrw.de/produkte/umwelt_klima/wald_forst/fernerkundung/baumartenklassifikation_nrw_EPSG25832_TIFF.zip
0 https://www.wms.nrw.de/umwelt/waldNRW?SERVICE=WMS&REQUEST=GetCapabilities

Tags

  • forst
  • opendata
  • waldinfo
  • wald
  • klassifizierung
  • baumarten
  • agri

Topics

  • ECON
  • ENVI
  • AGRI

Categories