Mapping secondary tropical forest and forest age from SPOT HRV data

Abstract

Accurate mapping of secondary forest and the age of these forests is critical to assess the carbon budget in tropical regions accurately. Using SPOT HRV (High Resolution Visible) data, techniques were developed and tested to discriminate primary forest, secondary forest and deforested areas on a study site in Rondonia, Brazil. Six co-registered SPOT HRV images (1986, 1988, 1989, 1991, 1992 and 1994) were used to create a time series of classified images of land cover (primary forest, secondary forest and deforested). These trajectories were used to identify secondary forest age classes relative to the most recent (1994) image. The resultant 1994 map of primary forest, secondary forest age classes and deforested areas served as ground reference data to establish training and testing sites. Several band 2 and 3 texture measurements were calculated using a 3 x 3 window to quantify canopy homogeneity. Neural networks and linear analysis techniques were tested for discriminating between primary forest, secondary forest and deforested pixels. The techniques were also employed to extract secondary forest age. A neural network using band 3 and a texture measure of band 2 and 3 from a single image (1994) discriminated primary forest, secondary forest (1 to >9 years) and deforested pixel with an average accuracy of 95%. The use of texture information increased the secondary forest discrimination accuracy 6.4% (from 83.5 to 89.9%). Spectral and textural information were also used to predict secondary forest age as a continuous variable. The neural network with the highest accuracy produced a RMSE (predicted network age versus actual secondary forest age) of 2.0 years with a coefficient of determination (predicted versus true) of 0.38. These results were significantly improved by using multitemporal information. The spectral and textural information from two images (1994 and 1989) were used to extract secondary forest information. The neural network results showed that 95.5% of the secondary forest pixels were correctly classified as secondary forest pixels (as opposed to 89.9% of the pixels using only the 1994 image). The RMSE and R2 accuracies in extracting secondary forest age as a continuous variable were 1.3 and 0.75, respectively.