1946 Ecological Applications, 14(6), 2004, pp. 1946–1962  2004 by the Ecological Society of America PREDICTING BIRD SPECIES RICHNESS USING REMOTE SENSING IN BOREAL AGRICULTURAL-FOREST MOSAICS MISKA LUOTO, 1,2,4 RAIMO VIRKKALA, 2 RISTO K. HEIKKINEN, 2 AND KALLE RAINIO 3 1 CSIRO Sustainable Ecosystems, Private Bag 5, P.O. Wembley 6914 WA, Australia 2 Research Programme for Biodiversity, Finnish Environment Institute, P.O. Box 140, FIN-00251 Helsinki, Finland 3 Department of Biology, University of Turku, FIN-20014 Turku, Finland Abstract. One of the main goals in nature conservation and land use planning is to identify areas important for biodiversity. One possible cost-effective surrogate for deriving appropriate estimates of spatial patterns of species richness is provided by predictive mod- eling based on remote sensing and topographic data. Using bird species richness data from a spatial grid system (105 squares of 0.25 km 2 within an area of 26.25 km 2 ), we tested the usefulness of Landsat TM satellite-based remote sensing and topographic data in bird species richness modeling in a boreal agricultural-forest mosaic in southwestern Finland. We built generalized linear models for the bird species richness and validated the accuracy of the models with an independent test area of 50 grid squares (12.5 km 2 ). We evaluated separately the modeling performance of habitat structure, habitat composition, topographical-moisture variables and all variables in the model-building and model-test areas. Areas of high observed and predicted bird species richness in the boreal agricultural- forest mosaic were mainly concentrated along river valleys in the grid squares with a high habitat diversity and steep topography. This landscape type also has the highest cover of habitats important for nature conservation: seminatural grasslands, deciduous forests, and watercourses. The covers of deciduous forests and seminatural grassland were included as explanatory variables of the bird species richness model, although they both cover only about 5% of the land in the study area. When the four models were evaluated by fitting them to the model test area, the explanatory power of the topography-moisture model decreased clearly, whereas the habitat-composition, habitat-structure, and all-variables mod- els were more rigorous. Finally, we extrapolated the models to the whole study area of 600 km 2 and produced bird species richness probability maps using GIS techniques. We conclude that, instead of scattered study plots in which birds are counted, predictive modeling requires large study areas where the variation within the whole landscape can be taken into account. A spatial grid system with several environmental variables derived from remote sensing data produces the most reliable data sets, which can be used when predicting species richness in other landscapes. Key words: agricultural-forest mosaic; biodiversity; bird species richness; boreal landscape; GIS; landscape management; predictive modeling; remote sensing; spatial grid system. INTRODUCTION One of the main goals in nature conservation and land-use planning is to identify and preserve areas im- portant for biodiversity. Unfortunately, data on spatial biodiversity patterns, particularly the distribution and richness of species within landscape mosaics, are often sparse or totally lacking and typically expensive to ac- quire (Margules and Austin 1991, Scott et al. 1993, Gaston 1996). National bird species records often have inadequate resolution (e.g., 10  10 km in bird atlas studies) for conservation or land-use planning. One po- tential way to overcome this difficulty is provided by predictive biodiversity models that can be applied over Manuscript received 28 May 2002; revised 23 October 2003; accepted 24 March 2004; final version received 13 April 2004. Corresponding Editor: C. A. Wessman. 4 Present address: Finnish Environment Institute, P.O. Box 140, FIN-00251 Helsinki, Finland. E-mail: miska.luoto@ymparisto.fi large areas (Griffiths et al. 1993, Stoms and Estes 1993, Dettmers and Bart 1999). The spatial prediction of biodiversity patterns from survey data has recently been recognized as a signifi- cant component of conservation planning (Gould 2000, Guisan 2002, Guisan and Zimmermann 2000, Austin 2002, Scott et al. 2002). Austin (1999) and Guisan and Zimmermann (2000) provide extensive reviews of the recent advances in statistical approaches that may be used for predictive modeling. One of their main con- clusions was that statistical modeling techniques such as generalized linear (GLM; McCullagh and Nelder 1989, Crawley 1993) and additive models (GAM; Yee and Mitchell 1991) appear to be increasingly popular as the statistical models to be used. This is due to the ability of GLM and GAM models to handle nonlinear relationships and different types of statistical distri- butions characterizing ecological data, and to the fact that they are technically closely related to traditional