UC Santa Barbara

Aerial imagery, as a useful tool for emergency management and response, still poses challenges to its effective (Joyce, Wright, et al. 2009). The challenges include information quality, accuracy, and the timeliness of information delivery (Joyce, Wright, et al. 2009). This research utilizes and updates the Remote Sensing Communication Model (RSCM) to configure time-sensitive remote-sensing-systems based on unpiloted aerial systems (UAS) and machine-learning-based damage detection. The UAS (DJI Mavic 1 and DJI Matrice 300) configurations’ navigation accuracies are tested with repeat station imaging (RSI) and traditional imaging. Utilizing real-time-kinematic corrections, the DJI Matrice 300 flown with the RSI method at six sites representing critical infrastructure at San Diego State University is shown to have more accurate repeated navigation to camera stations (0.16m vs. 0.21m for traditional imaging) and multi-date image pairs with the RSI method are shown to have better image co-registration mean absolute error (MAE) accuracy (2.2 pixels MAE vs. 4.4 pixels MAE). The Mask R-CNN machine-learning model (He et al. 2017) evaluated on bitemporal layer-stacked images detected damage (cracks) with a better mean intersection over union (mIoU) when used with the RSI images (83.7% mIoU vs. 72.5% mIoU). A customized convolutional neural network (CNN) and recurrent neural network (RNN) were evaluated using the RSI images, with the RNN having the higher accuracy (98.4% overall accuracy vs. 96.9% overall accuracy). Using the acquisition and newly demonstrated method of analyst capacity within the RSCM, the traditional imaging method and DJI Matrice 300 configuration had the highest timeliness capacity (1.02 x 107 bits m-1 s-1 vs. 7.40 x 106 bits m-1 s-1) of acquisition, and image co-registration based on the traditional imaging method had the highest co-registration analyst capacity (1.30 x 107 bits m-1 s-1 vs. 1.16 x 107 bits m-1 s-1). Of the three machine-learning models, the CNN had the highest analyst capacity and the RNN had the next highest (9.00 x 107 bits s-1 vs. 6.68 x 107 bits s-1).