The use of Unmanned Aerial Vehicles is feasible in cattle detection and counting, an essential activity in cattle farm management. The UAVs are alternatives to the traditional methods of gathering the animals in the pasture or taking them to the corral for this purpose, which can compromise herd health and productivity. They also replace the use of manned aerial vehicles, of high cost to cattle raisers. However, positioning the drone camera perpendicularly to the ground, as generally adopted when using unmanned vehicles, can bring limitations to animal monitoring, especially in extensive production systems. Researchers have shown that the inclined angle of the camera enlarges the area view from the pasture and reduces the quantity of required flights and harmful effects on the activity of animal detection. Scientists have adopted a computational architecture of deep neural networks to generate the models to be applied to the experiments. Article on preliminary results achieved was published in the journal Drones Oblique images and deep learning technologies, such as a computational neural network called convolutional, have proved promising for cattle detection and counting in the pasture through drone. This is what indicates preliminary results of studies described in the article Cattle Detection Using Oblique UAV Images, published in December by the journal Drones. The acronym UAV, Unmanned Aerial Vehicle. It is the first study uptaking the viability of the use of oblique images for cattle monitoring. The application of artificial intelligence algorithms to digital image processing and the progress of these technologies have shown the viability of this monitoring through unmanned aircraft. "However, the practical use is still a challenge due to the particular characteristics of this application, such as the need to track moving targets and the extensive areas that need to be covered in most cases," warn the researchers Jayme Garcia Arnal Barbedo and Luciano Vieira Koenigkan, from Embrapa Agricultural Informatics, and Patrícia Menezes Santos, from Embrapa Southeastern Livestock, authors of the article. The scientists then investigated the use of an inclined angle of the drone's camera to increase the area covered by the images, in order to minimize problems in tracking. Images captured under an oblique view, by enlarging the coverage, reduce the number of flights required for the activity, especially over large areas and decrease the harmful effects of animal movement and changes in environmental conditions. Studies employing UAVs for cattle monitoring almost always use images captured at the position perpendicular to the ground. In the process, the researchers applied a computational architecture of deep neural networks to generate the models applied to the experiments. Varied aspects were covered, such as ideal dimensions of the images, the effect of the distance between animals and the sensor, the effect of the classification error on the general detection process and the impact of physical obstacles on the accuracy of the model. Experimental results indicate that oblique images can be used successfully under certain conditions, but have practical and technical limitations that must be observed. These limitations refer to vision obstructions, the determination of the exact edges of the region considered in the images, geometric and color distortions, among others. Future investigations should include cost-benefit analysis to estimate the advantage from potential oblique images in comparison with the necessary measures to reduce practical obstacles. The experiments were performed with the purpose of detecting animals, an intermediate step for herd counting. Accurate count The practical part of the work was performed in extensive, intensive and integrated crop-livestock (ICL) systems at the Cachim farm, Embrapa Southeastern Livestock’s headquarters. For 2020, data collection in areas with trees and shrubs was planned, but the pandemic delayed the experiments. According to Patrícia Santos, trees, shrubs or even the height of the pasture can make it difficult to capture images. "The animal is hidden under the plant, hindering the count. To generate a model that corrects this, several images would be needed in areas with different trees and shrub plants, and heterogeneous shapes. Anything that can cover the image, even the height of the grass, should be considered. For example, a very high pasture can hide a calf, " she explains. There are many variables that the machine needs to learn in order that the cattle count is as accurate as possible. The scientist says that the role of Embrapa Southeastern Livestock is to help identify hindrances that may arise when the cattle farmer applies the tool day by day on the farm. The aim is to Indicate the real need of a potential user of this product, and estimate the acceptable margin of error. " A survey for inventory purposes does not allow errors. In the case of cattle counting for management, it can be a little more flexible”, Santos highlights. The researchers also point out that it is essential to expand knowledge about these techniques so that in the future the technology is successfully adopted in the field. “The results have been very good, but we still need more advances to be able to generate a technology that can be used by farmers or service providers. I believe we are on the right path”, assesses Barbedo. He estimates that monitoring with drones for automatic animal counting will happen in about two to three years. The methodology can also be used in the future to monitor animal health, such as the detection of diseases and anomalies and occurrences such as pregnancy. In this case the expectation is about five years. Horse monitoring The management of beef cattle farms under an extensive production system is challenging, especially considering that many of those farms have large areas with difficult ground access and insufficient communications infrastructure. Under those conditions, horse ground monitoring is a common practice. The alternative of areal herd inspection requires manned flights, expensive and subject to some risks. The farmer Renato Alves Pereira, owner of a property in the area of Mata Mineira, MG, says that the cattle count in his 830-hectare farm is done by two employees on horseback. And the cost is also high. He spends R$ 78,000 each year. The conference is performed weekly. This type of management usually requires the animals to be gathered in the corral. The change in routine by itself is already a stress factor for the animals. Both on the way to the corral and as it is handled, the cattle can stop feeding, drinking water, and resting. Research indicates that stressful situations cause direct impacts on the welfare and yield of the herd. A tool to count cattle by drone can be considered more rational. For Renato Pereira, who has been working with beef cattle farming for 40 years, the main advantages of such a technology would be the reduction of physical counting costs and the optimization of this process. "If the technology becomes operational, I have an interest in using it," he says. Computational models Studies aimed at cattle detection and counting through images captured by drones started in February 2019, valid for two years. They have financial support from the São Paulo State Research Foundation (Fapesp). About R$ 175,000 were invested, mainly in the acquisition of drones and equipment for image processing. In the studies conducted by Embrapa in partnership with the State University of Campinas (Unicamp) and the University Center of United Metropolitan Colleges (FMU), the team captured a great number of aerial imagery of Nelore breed animals (Bos indicus) and Canchim cattle – crossbreed between Charolês (Bos taurus) and Nelore, at the Embrapa farm. Afterward, it used the algorithm to classify them and extract the relevant information. In the case of cattle monitoring, the application includes animal detection and counting, specimen recognition, distance measurement between the cow and the calf, and determination of the feeding behavior. This information contained in the images is extracted through computational models of machine learning that use the concept of semantics and instance semantics, object detection, and heat mapping. Deep learning techniques are similar to those used in sites that request the user to identify images from crosswalks or traffic lights, for example, before accessing information considered of restricted use. That is, a need to practice the neural network with thousands of examples, teaching the computer to recognize the objects in an automatic way. Two articles, published in the journal Sensors, present the counting and detection models developed, in addition to preliminary results. The first, Counting Cattle in UAV Images-Dealing with Clustered Animals and Animal/Background Contrast Changes, is authored by Barbedo, Koenigkan , Patrícia Menezes, and the FMU researcher Andrea Roberto Bueno Ribeiro. The article proposes an algorithm that can provide accurate estimates for animal counting, even in difficult conditions, with the presence of grouped animals, changes in the contrast between them and previous ones, which is common due to the heterogeneity of cattle farms, and lighting variations. Some situations proved challenging, especially the lack of contrast between animals and the background, their movement, agglomeration of large animals, and the presence of calves. According to the authors, new solutions for animal tracking will be investigated in future experiments. The efforts should also be directed for the image capturing of another breed, possibly extending the algorithm applicability. Although the algorithm described in the article was developed in view of cattle counting, the methodology can be adapted to other applications such as the detection of ships or tents in refugee camps, among other applications. The second article on Sensors, A Study on the Detection of Cattle in UAV Images Using Deep Learning was written by Barbedo, Koenigkan, Thiago Teixeira Santos, also from Embrapa Agricultural Informatics, and Patrícia Menezes. In this study, the experiments involved 1,853 images containing 8,629 animal image samples. Thereby, 900 models were trained in convolutional neural networks, allowing a deep analysis of various aspects that impact cattle detection using aerial imagery captured by drones. The objectives were to determine the highest accuracy that could be achieved in the Cachim breed animal detection, as they are visually similar to the Nelore, as well as the ideal distance of ground sample for animal detection and a more precise architecture. Research leader is one of the most influential scientists in the world The researcher Jayme Garcia Arnal Barbedo (photo), from Embrapa Agricultural Informatics, is among the 100,000 most influential scientists in the world, according to the survey by Stanford University (USA) published in the journal Plos Biology. Another 16 Embrapa researchers are also part of the list, which considered the citations from the database Scopus, with more than 6,800 scientists, to assess their impact throughout their careers and in 2019. See the news article with more detail about this study here. The automatic recognition of diseases in plants using digital images was one of the first research lines I dedicated myself to at Embrapa. At the time, few research groups here and abroad approached the theme, so many of my studies were pioneering and worked as inspiration for researchers around the world", Barbedo recalls. "This confirms that the efforts spent in the last ten years have in fact contributed to the advance of agriculture and Brazilian science", he reinforces.