Object-based land cover maps reflecting different ecological habitats were generated for four study areas in the Torneträsk region, northern Sweden. Unmanned Aerial Vehicle (UAV) data was collected using an eBee fixed-wing mapping drone (senseFly, Switzerland) with a Sequoia multi-spectral (Parrot, France) camera sensor. The area was then segmented with all four bands (G,R,RE,NIR) at a resolution of 0.3 m using mean shift segmentation. Environmental data to inform the classification was extracted from early seasonal, peak seasonal and late seasonal UAV flights of 2018 using the spectral bands and the NDVI. Additional topographic layers were generated either from the UAV imagery or the Swedish national elevation data at 2 m resolution or laser data (Lantmäteriet). These additional layers were vegetation height, catchment slope, wetness index, relative slope position and slope. Based on the environmental data, the segments were then classified using a random forest machine learning classifier. Training areas were identified in the RGB UAV imagery and on field photographs. The classes describe common Arctic land cover types and reflect potential vole and lemming habitats. Accuracy was assessed using the out-of-bag error, Overall Accurcay and the Kappa index.

Accuracy assessment values of the land cover classifications.NF: OOB estimate of error rate: 3.22%, Accuracy : 0.67, Kappa : 0.60NT: OOB estimate of error rate: 5.67%, Accuracy : 0.58, Kappa : 0.50VJ: OOB estimate of error rate: 0.97%, Accuracy : 0.61, Kappa : 0.55KJ: OOB estimate of error rate: 0.32%, Accuracy : 0.70, Kappa : 0.65