Materials, Methods & Technologies, Volume 18, 2024

Rezumat: Accurate identification of individual wildlife is essential for effective species management and conservation. However, it becomes particularly challenging when unique features, like spot shapes and sizes, serve as the sole distinguishing characteristics, as in the case with Sika deer. This research aims to enhance the process of identifying individual Sika deer by employing a series of deep learning pipelines encompassing object detection, tracking, and automatic segmentation techniques facilitated by segment anything modules based on object detection. Through automatic segmentation contours are generated and various features including color channels, textures, and geometric attributes of the masked Sika deer, were extracted. Subsequently, Support Vector Machine (SVM) analysis was conducted to establish decision boundaries for individual Sika deer identification. We analyzed 2799 image datasets containing images of seven individual Sika deer captured by three camera traps deployed in farmland of Hokkaido, Japan over a span of 60 days. During our analysis, we observed multicollinearity (VIF > 10) within the datasets. To address this, Principal Component Analysis (PCA) was conducted on the extracted features, wherein PCA1 and PCA2, which collectively accounted for over 80% of the variance, were selected for subsequent Support Vector Machine (SVM) analysis. Utilizing the Radial Basis Function (RBF) kernel, which yielded a cross-validation score of 0.97, proved to be the most suitable choice for our research. Hyperparameter optimization, facilitated by the GridSearchCV library, determined the gamma value of 10 and C value as 100. The OneVsRest SVM classifier achieved an impressive overall accuracy of 0.99 for both training and testing datasets of Sika deer images. This study introduces a precise model for identifying individual Sika deer using video frame, aiding conservationists and researchers in effectively monitoring and protecting this species.