Fig. 1: Performance of deep neural networks (DNNs) trained with different cost functions and data sets in comparison to stacked species distribution models (SSDMs). | Nature Communications

Fig. 1: Performance of deep neural networks (DNNs) trained with different cost functions and data sets in comparison to stacked species distribution models (SSDMs).

From: Multispecies deep learning using citizen science data produces more informative plant community models

Fig. 1

a Ranks inferred from a left-out test set of n = 12,325 citizen science observations. b species-by-species AUC measuring the performance of predicting the distributions of 1345 individual species based on independent survey data. c site-by-site AUC measuring the performance of predicting the composition of communities at 1489 sites based on independent survey data. Note we weighted the taxonomically balanced citizen science test set (five observations per species) with the number of training observations per species to obtain scores that represent typical field observations (see methods). Boxes of DNNs are shown in gray and boxes of SSDMs are shown in white. Results of the DNN trained with high-resolution data are shown with dashed lines. Central lines in the boxplots indicate medians, boxes indicate interquartile ranges, and whiskers indicate 2.5 and 97.5 percentiles; NDCG refers to Normalized Discounted Cumulative Gain; CEL refers to Cross Entropy Loss; and lowres. and highres. represent low resolution and high resolution, respectively. Letters superimposed on boxes represent results of two-sided paired Wilcoxon tests, with no significant differences for models sharing the same letter and significant differences otherwise. The exact p-values can be found in Supplementary Table 3. Source data are provided as a Source Data file.

Back to article page