Classifying the Neotropical savannas of Belize using remote sensing and ground survey

Aim This paper evaluates a method of combining data from GPS ground survey with classifications of medium spatial resolution LANDSAT imagery to distinguish variations within Neotropical savannas and to characterize the boundaries between savanna areas and the associated gallery forests, seasonally dry forests and wetland communities. Location Rio Bravo Conservation Area, Orange Walk District, Belize, Central America. Methods Dry season LANDSAT data for 10 April 1993 and 9 March 2001 covering a conservation area of 240,000 acres (97,459 ha), were rectified to subpixel accuracy using ground control points positioned by GPS ground survey. The 1993 image was used to assess the accuracy with which the boundaries between the savanna matrix and gallery forests, high forests, wetlands and water bodies could be discriminated. The image was classified by a maximum likelihood (ML) classifier and the shapes and areas of forest and wetland classes were compared with an interpretation of these land cover types from 1: 24,000 aerial photography, mapped at 1: 50,000 scale in 1993. The 2001 image was used to assess whether different subtypes of savanna could be distinguished from LANDSAT data. This required the creation of a reference ('ground truth') data set for testing classifications of the image. One hundred and sixty sample patches (650 ha, distributed over an area of 7000 ha) of ten sub-types of savanna vegetation and associates identified using a physiognomic classification scheme, were delineated on the ground by GPS and divided into two subsets for training and testing. Continuous classifications of LANDSAT data covering the savannas were developed that estimated potential contributions from up to five sub-types of land cover (grassland, wetland, pine woodland, gallery forest and palmetto). The accuracy of each classification was assessed by comparison against ground data. An ML classification was also produced for the 2001 image using the same areas for training. This allowed a comparison of the relative accuracy of both continuous and Boolean ML methods for classifying savanna areas. Results The boundary between savannas and evergreen forests, gallery forests and open water in the study region could be delineated by the ML classifier to within 2 pixels (60 m) using LANDSAT imagery. However, the constituent sub-types within the savanna were poorly discriminated. Whilst the shape and extent of closed canopy forest, gallery forest, wetlands and water bodies agreed closely with the distributions interpreted from aerial photography, classes such as 'open pine savanna' or 'grassland' were only 45-65% accurate when tested against ground data. A continuous classification, estimating the proportions of three savanna vegetation subtypes (grassland, marshland and woodland) present in each pixel, correctly classified more of the ground data for these cover types than the comparable ML result. Proportional mixtures of the land cover estimated by the continuous classifier also compared realistically with the vegetation formations observed along ground transects. Main conclusions By using GPS, a ground survey of vegetation cover was accurately matched to remotely sensed imagery and the accuracy of delineating boundaries and classifying areas of savanna was assessed directly. This showed that ML classification techniques can reliably delineate the boundaries of savannas, but continuous classifiers more accurately and realistically represent the distribution of the subtypes comprising savanna land cover. By combining these ground survey and image classification methods, medium spatial resolution satellite sensor data can provide an affordable means for land managers to assess the nature, extent and distribution of savanna formations.