Species richness distributions underlie conservation strategies and the management of natural resources, but data issues including the lack of species surveys and geographical gaps are posing obstacles to evaluating biodiversity. In this study, we overcome the geographical gaps of data by using remote sensing data and deep learning (DL), which is rooted in artificial neural network theory.
To estimate plant species richness of the Korean Peninsula, we first estimated plant species richness in South Korea by stacking the probability-based predictions of 1,574 plant species from the Maxent species distribution model. Then, a multilayer perceptron (MLP) is used to develop a DL in North Korea where data is limited. Time series data of MODIS-driven NDVI and LAI, that may have high correlations with species richness, are exploited to develop a species richness model. Sixteen-day composite of NDVI and LAI images at 1km spatial resolution are used as input variables, and species richness computed in South Korea are used as a target variable. Due to the lack of ground truth in North Korea, training data in South Korea are divided into quadrants (NE, NW, SE, SW), and then k-fold cross validation is performed to determine the best hyperparameters.
Results/Conclusions
The mean AUC and mean Boyce index of 1,574 species’ species distribution models were 0.77 (± 0.11 SD) and 0.73 (± 0.18 SD), respectively. When we stacked the probability-based predictions of 1,574 species, the 1km cell with the highest number of species richness had 1,167 species and it was located in the middle of Jeju Island. Stacked species richness predicted that the north-eastern part of South Korea and Jeju Island contain the highest species richness areas, and the results corresponded with previous literature and other studies.
When comparing the results of the stacked species richness and the DL species richness model, we obtained 40.2 for RMSE, 0.95 for R2, and 0.99 for the slope of the regression model. While species distribution model is unable to assess species richness in areas where input variables are limited, remote sensing data were successfully incorporated into the state-of-the-art DL model quantitatively and qualitatively to evaluate species richness. By combing with advanced data technology, the application of remote sensing in biodiversity quantification can cope with information gaps in data-poor, but often biodiversity-rich areas.