Klassifizierung von Fernerkundungsdaten mittels logistischer Regression

Trippstadt

2008

5 Abb.,11 Lit. Ang.

Artikel aus einer ZeitschriftUnselbständiges Werk

14170

200149026

The feasibility of logistic regression models was determined for mapping of the following attributes: Occurrence of forest, Mixture of forest and the Existence of deadwood. Five test areas were chosen, representing more than 10 % of the area of Germany, in which models were fitted for both Landsat 5 and 7 data. In a small part of the fifth area data of the QuickBird system was analysed. The database of the German National Forest Inventory served as ground truth. Based on the Landsat 7 models forest distribution maps were generated. The regression models were tested for numeric stability by means of the k fold cross validation whereas the maps' accuracy was assessed using Orthophotos. The occurrence of forest was classified with a maximum accuracy of 95,4 % and mapped with 94,6 % based on Landsat 7. A rather low agreement of 83 % between ground truth and model Classification was achievable with the QuickBird-based models, therefore, no map was produced. The Classification of mixture classes was even less successful with overall accuracies between 75 and 83 %. Individual classes such as Mixed Forest ranged down to 39% of agreement, making the models only marginally better than Chance at Classification. The existence of deadwood was also predictable with logistic regression models based on Landsat 7 data. The variability of the results (overall agreements between 70 and almost 90 %) was greater than that of the attribute Forest occurrence.