COS 206-5 - Detecting Ecosystem Change From The Skies With Deep Learning

Background/Question/Methods

Worldwide, biodiversity is at extreme risk from anthropogenic change, including a human-driven warming climate and increased land use changes. Direct detection of biodiversity loss globally through field sampling at high resolution is infeasible given the time field collections take and the steep rates at which diversity is being lost. Recently, deep convolutional neural networks trained using high-resolution remote sensing data paired with citizen science observations have successfully predicted both individual plant species presence and overall plant communities at meters-level resolution, enabling the comprehensive detection of plant biodiversity at a fine scale. Here, we propose to use the same network to predict plant community distributions from 2012 to 2014 and then use the subsequent predictions to detect community changes in California. To do so, we developed a novel change metric that involves taking the euclidean distance of the difference between the predicted species presence probabilities acrossfrom the two timepoints. By aggregating the change in predicted presence across individual species, this unitless metric provides a community-level summary of an ecosystem’s plant community composition change.

Results/Conclusions

Since measuring community change at high resolution is still a largely unsolved question, we present a series of case studies from the model. First, we show that our model using the novel metric discussed above is able to detect ecosystem changes induced by the severe 2013 Rim Fire in the Sierra Nevada. Within the bounds of the fire, our metric is 9.7% higher than outside the bounds, overall predicting a significant increase in ecosystem change within areas affected by the fire (p< 2.2E-16). Second, our model is apparently able to detect wetland restoration in a seasonal estuary in Lassen County. Between 2012 and 2014, the Ash Creek Wildlife Area underwent a pond and plug restoration project to restore the threatened riparian habitat. Our model predicts significantly higher ecosystem change inside the park’s boundaries than outside (p< 2.2E-16). Furthermore, within the park itself, our model also predicts significantly higher change in areas within and around the restoration ponds than outside (p< 2.2E-16). Given the global availability of satellite imagery, this method serves as a powerful pilot for how citizen science observations can be paired with remote sensing data to predict biodiversity and its decline at continental scales.