The automatic analysis of satellite imagery has a wide range of applications within the field of urban planning, including fair distribution of resources, effective disaster response, updating of real-time maps and epidemiological vector-borne diseases control. Furthermore, it poses compelling technical challenges that even today are not completely solved. A system for multi-target building tracking using satellite images has been developed following the guidelines pro-posed in the SpaceNet 7 Multi-Temporal Urban Development Challenge and as a continuation of a previous theoretical exploration of the problem. The system was implemented by considering each individual block: a preprocessing stage, a neural network for semantic segmentation, and an algorithm for data assignment as a tracker for static targets. Even thoughthe dataset provides only images with moderate resolution and includes regions with high variability, crowded scenes, and a high number of targets, the model is able to segment correctly most of the buildings and maintain their identities along the sequence with a 62% of Intersection over Union (IoU). The system is able to locate correctly the buildings in the image and to determine accurately their borders with the exception of those too close to each other. Most importantly, the system reacts well to changes, which is an important factor of concern for urban planning purposes. The tracker reaches a MOTA of 0.647 and a F-score of 0.805 on the testing set. This repository has been developed as a tutored research project of the University of Bordeaux, the Autonomous University of Madrid and the Pázmány Catholic Peter's University under MIT license.

The data augmentation performed in the scope of this project can be grouped into two categories: color augmen-tations and geometrical augmentations. Both contribute to alleviate the small size of the dataset, but color augmentationsare specifically intended to add generalization ability regarding the high variability in the vegetation due to the geographical differences and seasonal changes. The processing outline for the two types of augmentations is the same: the whole dataset is traversed, and each augmentation is applied to an image with a probability of 20%

UNET with transfer learning was selected for the best results in three scenarios were implemented. In all of them the input size was set to 256x256, the binarization threshold to 0.6, the number of epochs to 100, and the initial learning rate to 1x10−4. However, learning rate decay was found to be useful in the three models. It is an strategy to train neural networks that improves both optimization and generalization, allowing a faster convergence of the loss function to a minimum and avoiding oscillation. To do so, the learning rate is decreased to 1x10−5 in epoch 30 and to 1x10−6 in epoch 65. Regarding the particular configurations, model 6 is the basic configuration (UNET+VGG16). Model 7 introduces a new loss function (hence forth mixed loss) proposed by one of the winners in a previous version of the challenge [21], whose implementation can be found in [23]. It corresponds to the sum of the focal loss and the dice loss. Since the ground truth of the dataset for which it was designed had several channels, it was computed as a weighted average over the different channels.For this case, being a binary class task, it corresponds to a simple sum. For tuning methods and other models trained please refer to docs.

The problem previously stated corresponds exactly to a data assignment problem. To solve it, the Hungarian Algorithm is used (sometimes referred as Munkres or Kuhn-Munkres algorithm). It is a strategy frequently employed in multiple object tracking systems [12] to find the optimal assignment between "workers" and "tasks" (in this case between the buildings detected in step t and those detected in step t+1) by minimizing a cost matrix. It has been proven that the algorithm solves the problem in a polynomial time, which is fast enough for the intended application.

The initial method is clearly suboptimal, especially considering the high number of targets in each image. The first and most relevant improvement is to narrow down the estimation of the cost by considering only the neighborhood of the footprint currently analysed. For this, a bounding box is generated for each footprint, considering its borders plus an offset of 20 pixels. The IOU for a pair of detections (in two consecutive frames) is computed only if the bounding boxes intercept. Otherwise, it will be set to 0. A second operation was done to the reduce the size of the matrices used. To explain this, let us consider the case when the footprints are located in the bottom right corner of the image. Originally, the IOU is computed with the entire mask (1024x1024) to preserve the spatial information. Nevertheless, it is possible to preserve such information using a smaller matrix size: first, the minimum x and y coordinates for the pair of detections is subtracted from every coordinate to reset the minimum as the origin. Then, the mask is created from the point (0,0) until the new x and y maximum for both cases. This allows to operate with considerably smaller matrices and, as a result, the processing time decreases.

An experiment was performed on a sample sequence of the dataset to measure the processing time and the effect of the optimizations. Only one time step was considered (i.e., the tracking is done for two frames only). The first image has 1500 targets and the second one 1520. The results can be seen in the report.