Unmanned Aerial Vehicles (UAVs) have shown great capability for disaster management due to their fast speed, automated deployment and low maintenance requirements. In recent years, disasters such as flooding are having increasingly damaging societal and environmental effects. To reduce their impact, real-time and reliable flood monitoring and prevention strategies are required. However, the limited battery life of small lightweight UAVs imposes efficient strategies to subsample the sensing field in this context. This paper proposes a novel solution to maximise the number of inspected flooded surface while keeping the travelled distance bounded. Our proposal solves the so-called continuous Travelling Salesman Problem (TSP), where the costs of travelling from one location to another depend not only on the distance, but also on the presence of water. To determine the optimal path between checkpoints, we employ the fast sweeping algorithm using a cost function defined from hyperspectral satellite maps identifying flooded regions. Preliminary results using MODIS flood maps show that our UAV planning strategy achieves a covered flooded surface approximately 3.33 times greater for the same travelled distance when compared to the conventional TSP solution. These results show new insights on the use of hyperspectral imagery acquired from UAVs to monitor water resources.
Infrared thermography (IRT, or thermal video) uses thermographic cameras to detect and record radiation in the longwavelength infrared range of the electromagnetic spectrum. It allows sensing environments beyond the visual perception limitations, and thus has been widely used in many civilian and military applications. Even though current thermal cameras are able to provide high resolution and bit-depth images, there are significant challenges to be addressed in specific applications such as poor contrast, low target signature resolution, etc. This paper addresses quality improvement in IRT images for object recognition. A systematic approach based on image bias correction and deep learning is proposed to increase target signature resolution and optimise the baseline quality of inputs for object recognition. Our main objective is to maximise the useful information on the object to be detected even when the number of pixels on target is adversely small. The experimental results show that our approach can significantly improve target resolution and thus helps making object recognition more efficient in automatic target detection/recognition systems (ATD/R).