Models predicting biota from environmental factors are useful for assessing environment. We compared three kinds of bird distribution models based on grid cell data: multiple regression models predicting guild species richness, logistic regression models predicting single species presence, and canonical discriminant models predicting species compositional types. Bird distribution data were obtained from bird censuses carried out in breeding and wintering seasons between 2000 and 2002. Bird species were classified into three guilds based on information of their ecology and habitat use. As environmental information, we used land cover map derived from satellite imagery taken by EOS Terra / ASTER sensor. We built the three kinds of models using the proportion of each land cover category in each cell. The results indicated that models predicting guild species richness are superior to those predicting presence of single species or species compositional types, in terms of versatility and possibility of extrapolation. Model performance was poorer for waterside species than for woodland species. It is presumably because of the problems in land cover classification.