Assessing Tree Cover in Agricultural Landscapes Using High-Resolution Aerial Imagery

Trees used in agroforestry practices, such as windbreaks, provide a variety of ecosystem benefits and are recognized globally as an important land use. However, efforts to inventory and monitor agroforestry land use have been sporadic, short-lived, or focused on small spatial extents. There are a variety of satellite-derived datasets that provide information about tree cover over broad spatial extents, but most are based on satellite sensors with resolutions too coarse to accurately observe narrow plantings of trees. We derived area estimates of land with tree cover in North Dakota and South Dakota from the National Land Cover Dataset, the Cropland Data Layer, MODIS Vegetative Continuous Fields, and a MODIS land cover product. We compared these image-based estimates to estimates based on in situ observations of forest land from the USDA Forest Service’s Forest Inventory and Analysis (FIA) program. Satellite-derived estimates of tree cover area differed from FIA forest land estimates by as much as 200,000 ha in both North Dakota and South Dakota. Image data from high resolution satellite sensors can detect small or narrow features, but prohibitively high data costs prevent their use for conducting national inventories. We used freely available, 1-m resolution imagery from the National Agriculture Imagery Program (NAIP) to map tree cover in Pembina County, North Dakota, USA. The approach used image segmentation and Random Forests, an ensemble classification tree algorithm. The Random Forests approach to mapping tree cover resulted in 84.8% agreement between model predictions and the out-of-bag sample. Based on the Gini index, texture attributes were more important predictors of tree cover than spatial or spectral attributes. Variability between flight lines in the NAIP imagery led to over-prediction of tree canopy in particular north/south swaths in the county. While future evaluation is required to develop an optimal training dataset to assess tree cover, the procedure shows promise for application over a broad spatial extent. ACknowledgmenTS We thank Susan Crocker, Dr. Randy Hamilton, and Dr. Kathleen Ward for thoughtful reviews of the early draft of this manuscript. We greatly appreciate the assistance with image interpretation by Cassandra Olson and Paul Sowers. The Journal of Terrestrial Observation | Volume 2 Number 1 (Winter 2010) Assessing Tree Cover in AgriCulTurAl lAndsCApes | 39 keYwoRdS image segmentation, Random Forests, NAIP, agroforestry