Mapping tropical dry forest succession using multiple criteria spectral mixture analysis

Tropical dry forests (TDFs) in the Americas are considered the first frontier of economic development with less than 1% of their total original coverage under protection. Accordingly, accurate estimates of their spatial extent, fragmentation, and degree of regeneration are critical in evaluating the success of current conservation policies. This study focused on a well-protected secondary TDF in Santa Rosa National Park (SRNP) Environmental Monitoring Super Site, Guanacaste, Costa Rica. We used spectral signature analysis of TDF ecosystem succession (early, intermediate, and late successional stages), and its intrinsic variability, to propose a new multiple criteria spectral mixture analysis (MCSMA) method on the shortwave infrared (SWIR) of HyMap image. Unlike most existing iterative mixture analysis (IMA) techniques, MCSMA tries to extract and make use of representative endmembers with spectral and spatial information. MCSMA then considers three criteria that influence the comparative importance of different endmember combinations (endmember models): root mean square error (RMSE); spatial distance (SD); and fraction consistency (FC), to create an evaluation framework to select a best-fit model. The spectral analysis demonstrated that TDFs have a high spectral variability as a result of biomass variability. By adopting two search strategies, the unmixing results showed that our new MCSMA approach had a better performance in root mean square error (early: 0.160/0.159; intermediate: 0.322/0.321; and late: 0.239/0.235); mean absolute error (early: 0.132/0.128; intermediate: 0.254/0.251; and late: 0.191/0.188); and systematic error (early: 0.045/0.055; intermediate: −0.211/−0.214; and late: 0.161/0.160), compared to the multiple endmember spectral mixture analysis (MESMA). This study highlights the importance of SWIR in differentiating successional stages in TDFs. The proposed MCSMA provides a more flexible and generalized means for the best-fit model determination than common IMA methods.