Wed, Aug 04, 2021:On Demand
Background/Question/Methods
A key challenge in quantifying tropical tree phenology lies in the fact that tropical forests contain hundreds of plant species with a wide variety and asynchrony of phenological patterns co-occurring within a landscape. Amazon dry season deciduousness has been shown to occur previously to a massive flushing in Central Amazonia with many consequences for ecosystem functioning. We still have a poor understanding about deciduous habits in the tropics, we do not know for example, how much and how long the forest experience deciduousness. Here we assess if deep learning algorithms applied to very high-resolution (VHR) images can effectively quantify changes in leaf cover, which we term deciduousness, in different landscapes across a seasonality gradient in the Amazon.
We use U-net convolutional neural network combined with VHR multispectral images (0.3 m) from WorldView-3 satellite to identify and segment deciduous tree canopies in three sites (one South and two East-Central Amazonia) along a seasonality gradient in the Amazon forest, in June/2016. We then quantify deciduousness cover (%) by empirically defining a spectral threshold of 90 (raw digital number) in the near-infrared (nir) band to delineate if the tree crown had a deciduous (leafless crowns – nir < 90 ) or semi-deciduous (crowns with exposed branches but presenting leaves – nir > 90) habit.
Results/Conclusions Our deep learning network segmentation was moderately-to-high accurate (> 80%) and Dice coefficients of above 0.63 for each landscape considered. We mapped 8148 deciduous tree crowns in the southern site, 5455 and 11028 in each landscape at East-Central Amazonia, representing a proportion of 32.5% (568.61 km2), 25.3% (254.17 km2), and 12% (2000 km2) of trees with different degrees of deciduousness within undisturbed forest in each landscape respectively. We found that nearly 70% deciduous trees mapped in southern Amazon can be considered “fully deciduous” (5754 trees), representing 19% of forest area, while 29% were classified as “semi-deciduous” (2394 trees; 6.3% of forest area). We found that most trees were fully deciduous in the two East-Central Amazonia landscapes, even though they presented lower proportions of deciduousness relative to the total forest area, with only 1% of semi-deciduous trees. Our work demonstrates large spatial variation in the degree of deciduousness among co-occurring tropical tree species within different landscapes in the Amazon, and a higher proportion of deciduousness cover than previously expected.Our results are a step towards to understanding the consequences of variation in tropical forest phenology for ecosystem processes and modeling.
Results/Conclusions Our deep learning network segmentation was moderately-to-high accurate (> 80%) and Dice coefficients of above 0.63 for each landscape considered. We mapped 8148 deciduous tree crowns in the southern site, 5455 and 11028 in each landscape at East-Central Amazonia, representing a proportion of 32.5% (568.61 km2), 25.3% (254.17 km2), and 12% (2000 km2) of trees with different degrees of deciduousness within undisturbed forest in each landscape respectively. We found that nearly 70% deciduous trees mapped in southern Amazon can be considered “fully deciduous” (5754 trees), representing 19% of forest area, while 29% were classified as “semi-deciduous” (2394 trees; 6.3% of forest area). We found that most trees were fully deciduous in the two East-Central Amazonia landscapes, even though they presented lower proportions of deciduousness relative to the total forest area, with only 1% of semi-deciduous trees. Our work demonstrates large spatial variation in the degree of deciduousness among co-occurring tropical tree species within different landscapes in the Amazon, and a higher proportion of deciduousness cover than previously expected.Our results are a step towards to understanding the consequences of variation in tropical forest phenology for ecosystem processes and modeling.