Detailed maps of tropical forests are within reach: Forest tree communities for Trinidad mapped with multiseason Landsat and Google Earth
Detailed maps of forest types are needed for research, REDD+ carbon accounting and biodiversity conservation, but spectral similarity among forest types; image cloud and scan-line gaps; and scarce vegetation plots complicate producing maps from satellite imagery. We used several key steps to mapping tropical forest habitats with cloudy Landsat to overcome these challenges. Our two main objectives were to test if 1) floristic classes of tropical forest trees can be mapped with multiseason, multidecade, gap-filled Landsat by judicious combination of field and remote sensing work; 2) synthetic multiseason Landsat imagery improves results. We also tested whether including the thermal band improved distinction of low-density urban areas.
We produced multidecade, multiseason gap-filled Landsat imagery with regression-tree normalization. Reference data came from field data, multiseason high-resolution imagery including like that on Google Earth, and Landsat archive imagery that was phenologically unique and not used in classification models. We used supervised decision tree classification.
We discovered that many floristic classes of tropical forest trees could be identified in high-resolution imagery or phenologically unique imagery by canopy structure, or by phenology on specific reference image dates, allowing extensive training data collection. In digital classifications, synthetic multiseason Landsat imagery significantly improved floristic class discrimination (by 14-21% for deciduous, 7-36% for semi-evergreen, and 3-11% for seasonal evergreen associations, and by 5-8% for secondary forest and woody agriculture). Imagery from climate extremes like severe drought increases accuracy the most. The seasonal spectral patterns in multiseason Landsat imagery have more spatial detail than most maps of environmental variables and can be more useful when mapping tropical forest tree communities with Landsat. The thermal band did not much improve low-density urban area mapping, but the approach we took to training data collection for these areas yielded excellent results.
Given a set of floristic tropical tree communities and general knowledge of their spatial distribution, features in multiseason and fine resolution reference imagery can allow discrimination among adjacent classes based on inundation or canopy deciduousness, flushing, flowering, or structure. The ability to distinguish classes in reference imagery permits the extensive training data collection needed to map floristic classes with noisy, gap-filled multiseason Landsat imagery. Forthcoming developments will make the key steps that we took to overcome the challenges of mapping detailed forest types in persistently cloudy regions accessible to nonspecialists.