Bering Sea | Mammals
Using georeferenced imagery data and Mask-RCNN neural network to detect and measure spotted seals Presenter: Alexey Altukhov , aaltukhov@gmail.com, North Pacific Wildlife Consulting Vladimir Burkanov , VBurkanov@gmail.com, National Marine Mammal Laboratory, AFSC, NMFS, NOAA, Seattle, USA; Kamchatka Branch of the Pacific G Denis Gaev , phototabmm@gmail.com The spotted seal ( Phoca largha ) size directly correlates with their age and can help to make inferences about the population structure and status. We used drone-based color imagery and convolutional neural network approach to assess the size composition of hauled out seals based on area outlined by each individual seal and to determine the proportion of various age cohorts in the population. To process drone imagery, we utilized photogrammetry software (Agisoft Metashape) with a set of developed python scripts to standardize mosaic creation and output as well as to apply procedures to reduce alignment/projection error. Overall, 100 mosaics were generated for 20 sites with a standard resolution of 0.5 cm per pixel and a standard projected output, UTM 58N. We initially trained our Mask RCNN model to detect individual seals on mosaics and outlined seal bodies on 111 tiles (section of mosaic 2048X2048 pxl); additional 43 tiles were used for training validation. The initial model was verified on tree independent datasets. For validation, we used a set of images similar to those that were used during training procedures, and it contained 618 individual animals. Control set consisted of images from sites that were not included in training and contained 165 individual animals. Raw drone imagery set contained 426 individual animals. False negative detection rate ranged 3% -7%. The initial algorithm outlined 83-86% of each individual seal body . The raw drone imagery had 34% of false negative detections, while algorithm outlined only 76% of each individual seal body . This suggested that more randomizations must be added to the model, thus we prepared an additional training dataset containing 429 image tiles that allowed us to improve model performance. Updated model false negative detection rate ranged 1% -3% while algorithm outlined 85-91% of each individual seal body in the validation and control image sets. Raw drone imagery processing output was also improved, but not greatly: false negative detection rate decreased to 27% while algorithm was able to mask only 79% of individual seal body. This suggests that raw imagery data still have high level of unaccounted variability and standardization of raw drone imagery must be performed prior to further analysis. Photogrammetry drone imagery processing is a one way of imagery data standardization that is proven to be effective in reducing variability in apparent seal sizes therefore reducing detection errors. Ivan Usatov , Usatov.ivan.Alex@gmail.com, Kronotsky Biosphere Nature Reserve Irina Trukhanova , irina_trukhanova@yahoo.com, North Pacific Wildlife Consulting LLC
Alaska Marine Science Symposium 2023 57
Made with FlippingBook flipbook maker