Watch Recording

Featured Presentations

Richie Carmichael (Biomark)

Drone Assisted Stream Habitat (DASH) Protocol: Establishing consistency and compatibility between UAS monitoring programs

Effective ecosystem management relies on accurate and timely evaluations of environmental status and trends, often equating to costly, time intensive survey efforts. Rapid advances in technology are constantly improving sampling methods, robust statistical inference, and thus cost and time efficiency. Perhaps one of the greatest steps in broad-scale habitat and wildlife monitoring has been advances in remote sensing technology. The Columbia River Basin is a major target for habitat restoration, working towards the de-listing of endangered Chinook salmon and steelhead. Multi-scale habitat characteristics are critical to understanding what defines quality habitat and where to focus restoration efforts. We developed the Drone Assisted Stream Habitat (DASH) protocol to collect data at the channel unit scale in a rapid manner, which is then paired with multispectral imagery collected via drone. Thanks to the time and cost efficiency of drone surveys, this approach can be easily applied to larger scales (tributary, watershed) with minimal additional on-the-ground sampling. Furthermore, we have developed tools that automate the post-processing of drone imagery, substantially increasing the cost efficiency and ease of post-processing. This approach allows for the pairing of fish and habitat data at multiple spatial scales ranging from the watershed to the channel-unit. These data can then be used to populate fish-habitat models, such as quantile regression forest (QRF) capacity models at any desired scale. In the Lemhi River, Salmon ID, we have applied DASH and QRF to define quality juvenile Chinook salmon and steelhead habitat, identify current capacity limitations, and monitor the effectiveness of restoration actions. Taken together, the two-pronged DASH and QRF approach is a comparatively inexpensive tool to prioritize, direct, and monitor habitat restoration in near real-time.

Sarah Hoffmann (Biomark)

Machine learning applications for conservation

Effective management of imperiled species, and the habitats they rely on, depend largely on accurate and timely environmental sampling. These data collection techniques are often costly, time intensive, or impossible due to inaccessible habitats. Advances in remote sensing techniques, especially the availability of unmanned aerial systems (UAS, drones), have vastly improved the efficiency of data collection; thus, the new bottleneck occurs at the data processing step. Image processing (orthorectification, alignment, photogrammetry, data extraction, and analysis), data storage, and computing requirements are all documented barriers to entry for remote sensing applications in the conservation world. To address this, we are employing machine learning techniques to automate the processing of imagery and extraction of data. Multi-spectral drone imagery is calibrated to generate absolute values of reflectance and eliminate minute differences between sensor capture timing as the drone is moving. A contrast limited adaptive histogram equalization (CLAHE) is applied to increase contrast and definition, thereby improving application of classifiers. We employed both a pixel-based random forest classifier as well as object-based detection in order to classify water, bare earth, vegetation, and woody debris. We are currently working to refine these classifiers in order to extract increased detail at the habitat level. Within the confines of a mask regional convolutional neural network model (rCNN), we are able train a variety of datasets, including the ability to detect and track marine megafauna throughout the southeast Florida coast. Given the proper training data, this neural network classifier is seemingly applicable to a wide variety of ecosystems and species. Our goal is to develop tools that provide real-time, actionable intelligence to drive the recovery of imperiled species.