Understanding climate change effects in alpine meadows: harnessing the power of data science, and remote sensing

Abstract

Montane ecosystems are particularly sensitive to climate warming. This is because Alpine wildflowers that are dominant in these ecosystems are prone to local level extinction (as suitable habitats disappear), as spring and summer temperatures increase. This can either occur when trees (previously limited by climate) invade Alpine meadows or Alpine plants encounter limited habitats at upper range limits into which to expand following warming. Climate-change induced shifts in Alpine wildflower phenology, observed for numerous species, can also have negative impacts by causing flowers to be exposed to damaging climatic conditions (e.g., early frosts) or causing pollinators to become mismatched with flowering. To anticipate these changes in natural areas, we need tools with which to monitor climatic factors influencing Alpine wildflower habitats as well as climatic impacts on Alpine Wildflowers. Unfortunately, the effects of warming, both on the climatic factors influencing Alpine wildflower habitats and phenology are difficult to quantify at the spatial and temporal scales needed. For example, many studies in Alpine habitats rely on climate stations to detect shifts in warming or snow melt, but these stations are generally few in number, and therefore fail to capture the spatial variability necessary to understand snowmelt patterns. Discernible phases of Alpine plant phenology have been collected by routine field measurements during the growing season, some using in-situ observations and others aided by citizen science efforts (e.g., National Phenology Network). Although such investigations have been instrumental in quantifying phenological shifts, they are challenged by the fact that limited resources often make it difficult to gather observations in remote locations, over large spatial scales and at sufficient frequency to capture phenological transitions. Recent technological developments in imaging, from the exploding and ubiquitous use of smartphones, increasingly improved UAV devices (drones), and satellites offer great promise in meeting this challenge. Of course, satellite-based imagery (e.g., MODIS, Landsat 8 and Sentinel) have been key for looking at spectral signatures to detect changes in plant phenology in other systems, but its main limitation in Alpine systems has been its coarse spatial resolution (30 - 1000m) and low temporal frequency (biweekly to once a month). This is insufficient to monitor Alpine systems, which occupy relatively narrow geographic areas (between treeline and rock / ice), and where seasonal dynamics are constrained to very short summer seasons (e.g., 3 months). The advent of CubeSat imagery that provides daily 3 - 5m resolution imagery (e.g. Planet, Digital Globe, etc.) has opened the possibility of detecting snow dynamics and multiple phenological changes (e.g., green-up, flowering) at much higher spatiotemporal accuracy and sustainability. Simultaneous advances in quantitative approaches to analyzing images (e.g., machine learning), paired with an increasingly high volume of images available from the general public, similarly provides new opportunities to detect Alpine flowering at increasingly small spatial scales. In this dissertation, I explore the promise and challenge of using images to better understand Alpine systems, particularly in terms of their phenology. In particular, I wished to develop tools that both capture the environmental drivers of that phenology (in our system and many Alpine systems, snow) as well allow us to monitor the progression of phenology at the landscape and individual species level. Specifically, I first (in Chapter one) sought to improve high-resolution (meter-scale) mapping of snow-covered areas in complex and forested terrain. Such mapping is critical for understanding the responses of montane flowering species to climate change. Here, myself and my colleagues used high-resolution imagery from PlanetScope to derive snow-covered areas and build on a previous work that demonstrated snow cover area mapping using only the PlanetScope 4-band (Red, Green, Blue and NIR) reflectance with convolutional neural networks (CNN) based machine learning (ML) approach. We augmented the existing CNN model with additional input data including vegetation metrics (Normalized Difference Vegetation Index) and DEM-derived metrics (elevation, slope and aspect) to improve SCA mapping in forested and open terrain (like alpine meadows) and showed that an augmented model that used the Normalized Difference Vegetation Index (NDVI) along with visible (red, green, and blue) and NIR bands was the best performing model. The NDVI based model with an F-score of 0.89 (Gunnison) and 0.93 (Engadin) was found to be 4% and 2% better than when using canopy height and terrain-derived measures at Gunnison, respectively. We examined the model’s performance in forested areas using three forest canopy quantification metrics and found that augmented models better identified snow in canopy edges and open areas but still underpredicted snow cover under forest canopies. The improved high-resolution snow maps in forested environments can support studies involving climate change effects on a broad range of phenomena in mountain ecosystems, including Alpine wildflower phenology as well as evaluations of hydrological impacts in snow-dominated river basins. In chapter two, my collaborators and I evaluated the use of hyperspectral CubeSat imagery to develop models for detecting peak flowering phenology - hypothesizing that reflectance in visible and NIR bands from CubeSat imagery could be used to distinguish flowering pixel events. I examined Alpine wildflower meadows at Mt. Rainier National Park (MORA) using Random Forest (RF) classification from high-resolution (3-m PlanetScope from Planet Labs, Inc., San Francisco, CA, USA) and moderate resolution (10-m Sentinel and 30-m Landsat) imagery. On-the ground citizen science data collected by the MeadoWatch program was used to validate these models. Peak flowering delineation using 3-m resolution PlanetScope imagery resulted in an accuracy of 70% (Cohen’s kappa = 0.25), whereas, combining with other sources of imagery (10-m Sentinel and 30-m Landsat) improved the delineation of peak flowering (accuracy = 77%, Cohen’s kappa = 0.39). This approach was also able to identify the timing of peak flowering in a particularly warm year (2015), despite being calibrated on normal climate years. In all, this work suggests that PlanetScope imagery holds promise for detecting community-level Alpine wildflower phenology, and is therefore particularly useful in global change ecology, where temporal frequency is especially important. Additionally, combining imagery may provide a new approach for cross-calibrating sensors to account for radiometric irregularity inherent in fine resolution PlanetScope imagery. In Chapter three, I explore the use of machine learning in crowd sourced camera images to quantify flowering species richness in the same meadows. I used three deep learning techniques (Mask R-CNN, RetinaNet and YOLO) to detect wildflower species in images taken during two flowering seasons and found that deep learning techniques can detect multiple species, providing information on flowering richness in photographed meadows. Two-stage detector Mask R-CNN was more accurate than single-stage detectors like RetinaNet and YOLO, with the Mask R-CNN network performing best overall with mean average precision (mAP) of 0.67 followed by RetinaNet (0.5) and YOLO (0.4). The results indicate higher richness just above the tree line for most of the species, which is comparable with patterns found using field studies. We emphasize the application of these approaches in many other ecological questions that benefit from automated flower detection, like studies of flowering phenology or floral resources for pollinators, and that this approach can therefore complement a wide range of existing ecological approaches, e.g., community science, experiments, etc. The development of this approach for wildflower phenology predictions provides new possibilities to monitor climate change effects on flowering communities at high temporal scales in locations where regular access is challenging. In combination, I took advantage of rapid technological and quantitative developments to build tools that allow ecologists to better monitor climate change impacts in sensitive Alpine wildflower meadows. However, I believe these tools can widely be applied in any ecologically sensitive system where processes of interest are rapid and can be captured in images, but where regular access by scientists to monitor these processes is expensive or difficult (e.g., remote islands, high canopies in trees, etc.).