Automatic vehicle identification from Google Earth satellite images based on single shot deep learning neural networks for object detection

Road and urban transport network facilitate our daily life for optimal routing. In the road network, traffic management is one of the main challenges of managers. In this regard, the first step is to estimate the density of cars at the level of the urban road network. Estimating the number of cars or the level of occupancy of cars in the whole city, taking into account less time and cost, is only possible with satellite images. In this regard, in this research, satellite images with high spatial resolution available and downloadable from the Google Earth system have been used. To identify the position of the cars, the single shot deep learning method with RetinaNet architecture and based on residual neural networks with the number of layers 18, 34 and 50 have been used. For the training data, the position of the cars is marked with bounding boxes and then satellite images with dimensions of 128 x 128 pixels and 64 pixels pitch are cut. Of the total training data, 80% have been used for training and 20% for validation. The models were trained in 50 epochs and with an average accuracy of 0.7. Satellite images containing more than 15000 cars was used to evaluate the trained models. The parameter of the possibility of overlapping of the non-maximum suppression method was applied equal 25%. The final result shows that the use of the proposed model in the identification of cars has a good accuracy. The RetinaNet detector model based on the residual deep learning network with 50 layers has performed best in terms of average accuracy with 0.87, precision with 0.7, recall with 0.99 and F1-score with 0.82. The main challenge of the proposed models is in areas with high car density, which reduces the possibility of accurately detecting the number of cars due to the size of the ground sampling distance of satellite images, but it estimates the occupancy level better.