Lesser snow geese (Chen caerulescens caerulescens) of the Western Canadian Arctic Population feed intensively for 2-4 weeks on the coastal plain of the Beaufort Sea in Canada and Alaska at the beginning of their autumn migration. Petroleum leasing proposed for the Alaskan portion of the staging area on the Arctic National Wildlife Refuge (ANWR) could affect staging habitats and their use by geese. Therefore we studied availability, distribution, and use by snow geese of tall and russett cotton-grass (Eriophorum angustifolium and E. russeolum, respectively) feeding habitats on the ANWR. We studied selection of feeding habitats at 3 spatial scales (feeding sites [0.06 m²], feeding patches [ca. 100 m²], and feeding areas [>1 ha]) during 1990-93. We used logistic regression analysis to discriminate differences in soil moisture and vegetation between 1,548 feeding sites where snow geese exploited individual cotton-grass plants and 1,143 unexploited sites at 61 feeding patches in 1990. Feeding likelihood increased with greater soil moisture and decreased where nonforage species were present. We tested the logistic regression model in 1991 by releasing human-imprinted snow geese into 4 10 × 20-m enclosed plots where plant communities had been mapped, habitats sampled, and feeding probabilities calculated. Geese selected more feeding sites per square meter in areas of predicted high quality feeding habitat (feeding probability ≥ 0.6) than in medium (feeding probability = 0.3-0.59) or poor (feeding probability < 0.3) quality habitat (P < 0.0001). Geese increasingly used medium quality areas and spent more time feeding as trials progressed and forage was presumably reduced in high quality habitats. We examined relationships between underground biomass of plants, feeding probability, and surface microrelief at 474 0.06- m² sites in 20 thermokarst pits in 1992. Feeding probability was correlated with the percentage of underground biomass composed of cotton-grass (r = 0.56). Feeding probability and relative availability of cotton-grass forage were highest in flooded soils along the ecotone of flooded and upland habitats. In 1992, we also used the logistic regression model to estimate availability of high quality feeding sites on 192 80 × 90-m plots that were randomly located on 24 study areas. A mean of 1.6% of the area sampled in each plot was classified as high quality feeding habitat at 23 of the study areas. Relative availability of high quality sites was highest in troughs, thermokarst pits, and water tracks because saturated soils in those microreliefs were dominated by cotton-grass. Relative availability of high quality sites was lower in saturated soils of basins (low-centered polygons, wet meadows, and strangmoor) because that microrelief was dominated by Carex spp. Most (63%) of the saturated area on the ANWR coastal plain was in basins. We examined distribution of feeding patches relative to microrelief in 49 snow goose feeding areas in 1993. Only 2.5% of the tundra in each feeding area was exploited by snow geese. Snow geese preferentially fed in thermokarst pits, water tracks, and troughs, and avoided basins and uplands. Feeding areas had more thermokarst pit but less basin microrelief than adjacent randomly-selected areas. Thermokarst pits and water tracks occurred most frequently in regions of the coastal plain where geese were observed most often during aerial surveys (1982-93). Microrelief influenced selection of feeding patches and feeding areas and may have affected snow goose distribution on the ANWR. Potential feeding patches were widely distributed but composed a small percentage (≤2.5%) of the tundra landscape and were highly interspersed with less suitable habitat. The Western Canadian Arctic Population probably used a large staging area on the Beaufort Sea coastal plain because snow geese exploited a spatially and temporally heterogeneous resource.