Search

Maintaining vegetation cover in cities is a key component to keep cities safe and resilient. The monitoring of trees is usually done with LiDAR data or multi and hyperspectral images. In this sense, remote sensing RGB images are presented as a cheaper and easier processing solution. Here, we proposed to evaluate deep learning-based methods combined with high-resolution RGB images to detect single-trees in the urban environment. Three state-of-the-art methods are tested: Faster-RCNN, RetinaNet, and ATSS. A total of 220 images were used, in which we manually labeled 3382 trees. For the proposal task, our findings show that ATSS is 3% more accurate than Faster-RCNN and 4% than RetinaNet. However, in a qualitative inspection, Faster-RCNN and RetinaNet seems to be better at this task. Our findings shows the need of further research for developing suitable tools for urban tree detection. This tools can help cities top achieve a more sustainable and resilient environment especially to face climate change.

Made available in DSpace on 2022-04-29T08:30:21Z (GMT). No. of bitstreams: 0 Previous issue date: 2021-07-01 Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) Urban forests contribute to maintaining livability and increase the resilience of cities in the face of population growth and climate change. Information about the geographical distribution of individual trees is essential for the proper management of these systems. RGB high-resolution aerial images have emerged as a cheap and efficient source of data, although detecting and mapping single trees in an urban environment is a challenging task. Thus, we propose the evaluation of novel methods for single tree crown detection, as most of these methods have not been investigated in remote sensing applications. A total of 21 methods were investigated, including anchor-based (one and two-stage) and anchor-free state-of-the-art deep-learning methods. We used two orthoimages divided into 220 non-overlapping patches of 512 × 512 pixels with a ground sample distance (GSD) of 10 cm. The orthoimages were manually annotated, and 3382 single tree crowns were identified as the ground-truth. Our findings show that the anchor-free detectors achieved the best average performance with an AP50 of 0.686. We observed that the two-stage anchor-based and anchor-free methods showed better performance for this task, emphasizing the FSAF, Double Heads, CARAFE, ATSS, and FoveaBox models. RetinaNet, which is currently commonly applied in remote sensing, did not show satisfactory performance, and Faster R-CNN had lower results than the best methods but with no statistically significant difference. Our findings contribute to a better understanding of the performance of novel deep-learning methods in remote sensing applications and could be used as an indicator of the most suitable methods in such applications. Faculty of Engineering Architecture and Urbanism and Geography Federal University of Mato Grosso do Sul Faculty of Computer Science Federal University of Mato Grosso do Sul Department of Cartography São Paulo State University (UNESP) Computing Science and Mathematics Division Faculty of Natural Sciences University of Stirling Department of Cartography São Paulo State University (UNESP) CNPq: 303559/2019-5 CNPq: 304052/2019-1 CNPq: 433783/2018-4 CAPES: 88881.311850/2018-01

Recently, methods based on Convolutional Neural Networks (CNN) achieved impressive success in semantic segmentation tasks. However, challenges such as the class imbalance and the uncertainty in the pixel-labeling process are not completely addressed. As such, we present a new approach that calculates a weight for each pixel considering its class and uncertainty during the labeling process. The pixel-wise weights are used during training to increase or decrease the importance of the pixels. Experimental results using different datasets like aerial, terrestrial, and ultrasound images show that the proposed approach leads to significant improvements in three challenging segmentation tasks in comparison to baseline methods. It was also proved to be more invariant to noise. The approach presented here may be used within a wide range of semantic segmentation methods to improve their robustness.

Deep Neural Networks (DNNs) learn representation from data with an impressive capability, and brought important breakthroughs for processing images, time-series, natural language, audio, video, and many others. In the remote sensing field, surveys and literature revisions specifically involving DNNs algorithms' applications have been conducted in an attempt to summarize the amount of information produced in its subfields. Recently, Unmanned Aerial Vehicles (UAV) based applications have dominated aerial sensing research. However, a literature revision that combines both "deep learning" and "UAV remote sensing" thematics has not yet been conducted. The motivation for our work was to present a comprehensive review of the fundamentals of Deep Learning (DL) applied in UAV-based imagery. We focused mainly on describing classification and regression techniques used in recent applications with UAV-acquired data. For that, a total of 232 papers published in international scientific journal databases was examined. We gathered the published material and evaluated their characteristics regarding application, sensor, and technique used. We relate how DL presents promising results and has the potential for processing tasks associated with UAV-based image data. Lastly, we project future perspectives, commentating on prominent DL paths to be explored in the UAV remote sensing field. Our revision consists of a friendly-approach to introduce, commentate, and summarize the state-of-the-art in UAV-based image applications with DNNs algorithms in diverse subfields of remote sensing, grouping it in the environmental, urban, and agricultural contexts., 27 pages, 10 figures

In recent years, many agriculture-related problems have been evaluated with the integration of artificial intelligence techniques and remote sensing systems. Specifically, in fruit detection problems, several recent works were developed using Deep Learning (DL) methods applied in images acquired in different acquisition levels. However, the increasing use of anti-hail plastic net cover in commercial orchards highlights the importance of terrestrial remote sensing systems. Apples are one of the most highly-challenging fruits to be detected in images, mainly because of the target occlusion problem occurrence. Additionally, the introduction of high-density apple tree orchards makes the identification of single fruits a real challenge. To support farmers to detect apple fruits efficiently, this paper presents an approach based on the Adaptive Training Sample Selection (ATSS) deep learning method applied to close-range and low-cost terrestrial RGB images. The correct identification supports apple production forecasting and gives local producers a better idea of forthcoming management practices. The main advantage of the ATSS method is that only the center point of the objects is labeled, which is much more practicable and realistic than bounding-box annotations in heavily dense fruit orchards. Additionally, we evaluated other object detection methods such as RetinaNet, Libra Regions with Convolutional Neural Network (R-CNN), Cascade R-CNN, Faster R-CNN, Feature Selective Anchor-Free (FSAF), and High-Resolution Network (HRNet). The study area is a highly-dense apple orchard consisting of Fuji Suprema apple fruits (Malus domestica Borkh) located in a smallholder farm in the state of Santa Catarina (southern Brazil). A total of 398 terrestrial images were taken nearly perpendicularly in front of the trees by a professional camera, assuring both a good vertical coverage of the apple trees in terms of heights and overlapping between picture frames. After, the high-resolution RGB images were divided into several patches for helping the detection of small and/or occluded apples. A total of 3119, 840, and 2010 patches were used for training, validation, and testing, respectively. Moreover, the proposed method’s generalization capability was assessed by applying simulated image corruptions to the test set images with different severity levels, including noise, blurs, weather, and digital processing. Experiments were also conducted by varying the bounding box size (80, 100, 120, 140, 160, and 180 pixels) in the image original for the proposed approach. Our results showed that the ATSS-based method slightly outperformed all other deep learning methods, between 2.4% and 0.3%. Also, we verified that the best result was obtained with a bounding box size of 160 × 160 pixels. The proposed method was robust regarding most of the corruption, except for snow, frost, and fog weather conditions. Finally, a benchmark of the reported dataset is also generated and publicly available.

Mapping utility poles using side-view images acquired with car-mounted cameras is a time-consuming task, mainly in larger areas due to the need for street-by-street surveying. Aerial images cover larger areas and can be feasible alternatives although the detection and mapping of the utility poles in urban environments using top-view images is challenging. Thus, we propose the use of Adaptive Training Sample Selection (ATSS) for detecting utility poles in urban areas since it is a novel method and has not yet investigated in remote sensing applications. Here, we compared ATSS with Faster Region-based Convolutional Neural Networks (Faster R-CNN) and Focal Loss for Dense Object Detection (RetinaNet ), currently used in remote sensing applications, to assess the performance of the proposed methodology. We used 99,473 patches of 256 × 256 pixels with ground sample distance (GSD) of 10 cm. The patches were divided into training, validation and test datasets in approximate proportions of 60%, 20% and 20%, respectively. As the utility pole labels are point coordinates and the object detection methods require a bounding box, we assessed the influence of the bounding box size on the ATSS method by varying the dimensions from 30×30 to 70×70 pixels. For the proposal task, our findings show that ATSS is, on average, 5% more accurate than Faster R-CNN and RetinaNet. For a bounding box size of 40×40, we achieved Average Precision with intersection over union of 50% (AP50) of 0.913 for ATSS, 0.875 for Faster R-CNN and 0.874 for RetinaNet. Regarding the influence of the bounding box size on ATSS, our results indicate that the AP50 is about 6.5% higher for 60×60 compared to 30×30. For AP75, this margin reaches 23.1% in favor of the 60×60 bounding box size. In terms of computational costs, all the methods tested remain at the same level, with an average processing time around of 0.048 s per patch. Our findings show that ATSS outperforms other methodologies and is suitable for developing operation tools that can automatically detect and map utility poles.

As key-components of the urban-drainage system, storm-drains and manholes are essential to the hydrological modeling of urban basins. Accurately mapping of these objects can help to improve the storm-drain systems for the prevention and mitigation of urban floods. Novel Deep Learning (DL) methods have been proposed to aid the mapping of these urban features. The main aim of this paper is to evaluate the state-of-the-art object detection method RetinaNet to identify storm-drain and manhole in urban areas in street-level RGB images. The experimental assessment was performed using 297 mobile mapping images captured in 2019 in the streets in six regions in Campo Grande city, located in Mato Grosso do Sul state, Brazil. Two configurations of training, validation, and test images were considered. ResNet-50 and ResNet-101 were adopted in the experimental assessment as the two distinct feature extractor networks (i.e., backbones) for the RetinaNet method. The results were compared with the Faster R-CNN method. The results showed a higher detection accuracy when using RetinaNet with ResNet-50. In conclusion, the assessed DL method is adequate to detect storm-drain and manhole from mobile mapping RGB images, outperforming the Faster R-CNN method. The labeled dataset used in this study is available for future research.

Scene recognition is an important and challenging problem in computer vision. One of the most used scene recognition methods is the bag-of-visual words. Despite the interesting results, this approach does not capture the detail richness of spatial information of the visual words on the image. In this paper, we propose a new method to describe the visual words using the fractal dimension. Our method estimates the fractal dimension of each visual word on image through box-counting method. The fractal dimension is capable of providing complex and spatial information of the visual words in a simple and efficient way. We validate our method on three well-known scene and object datasets, and the experimental results reveal that our method leads to highly discriminative features of the visual words. In addition, the proposed method has achieved competitive results compared to popular methods in scene classification.

In Multi-Label Stream Classification (MLSC) examples arriving in a stream can be simultaneously classified into multiple classes. This is a very challenging task, especially considering that new classes can emerge during the stream (Concept Evolution), and known classes can change over time (Concept Drift). In real situations, these characteristics come together with a scenario with Infinitely Delayed Labels, where we can never access the true class labels of the examples to update classifiers. In order to overcome these issues, this paper proposes a new method called MultI-label learNing Algorithm for Data Streams with Binary Relevance transformation (MINAS-BR). Our proposal uses a new Novelty Detection (ND) procedure to detect concept evolution and concept drift, being updated in an unsupervised fashion. We also propose a new methodology to evaluate MLSC methods in scenarios with Infinitely Delayed Labels. Experiments over synthetic data sets attested the potential of MINAS-BR, which was able to adapt to different concept drift and concept evolution scenarios, obtaining superior or competitive performances in comparison to literature baselines.

In multi-label classification problems an example can be simultaneously classified into more than one class. This is also a challenging task in Data Streams (DS) classification, where unbounded and non-stationary distributed multi-label data contain multiple concepts that drift at different rates and patterns. In addition, the true labels of the examples may never become available and updating classification models in a supervised fashion is unfeasible. In this paper, we propose a Multi-Label Stream Classification (MLSC) method applying a Novelty Detection (ND) procedure task to update the classification model detecting any new patterns in the examples, which differ in some aspects from observed patterns, in an unsupervised fashion without any external feedback. Although ND is suitable for multi-class stream classification, it is still a not well-investigated task for multi-label problems. We improve a initial work proposed in [1] and extended it with a new Pruned Sets (PS) transformation strategy. The experiments showed that our method presents competitive performances over data sets with different concept drifts, and outperform, in some aspects, the baseline methods.

Approaches that combine streaming algorithms and distributed computing have potential to deal with voluminous and high-speed data streams. Considering the data stream clustering task, also an important issue needs to be addressed, estimate the number of clusters dynamically, since it may vary due to concept drift. This work proposes three evolutionary-based algorithms to overcome these requirements. They are based in the discretized stream model, where the sequential batches of objects are distributed and processed in a parallel way using the MapReduce model. The proposed algorithms achieve superior experimental results, either in quality and processing time, overcoming the state-of-the-art.

The constant increasing of the generated data in real time has been creating new challenges for machine learning tasks and for one of their main branches: data clustering. The scenario of big data stream (high-speed data stream) has become reality. In order to deal with this scenario, new approaches are required. In this work, we present new techniques based on three clustering fields: Data Stream, MapReduce and automatic estimation of k from data. The goal is to cluster a high-speed data stream with varying number of clusters. Two scalable algorithms are proposed, based on centralized data stream algorithms. The first is based on the StreamKM++ and the second based on the F-EAC, an evolutionary algorithm used for batch clustering purpose. Results achieved the same high quality as the original centralized versions of the algorithms. The proposed techniques can be used instead the centralized ones when the velocity/volume is so high to fit in a centralized system.

Orbital emphysema can be a rare and severe postoperative complication. Fast and critical diagnosis is essential to avoid permanent visual loss to the patient. Therefore, a case of a male patient victim of a motorcycle accident with a postoperative complication of orbital emphysema has been presented. The immediate diagnosis and treatment of the orbital emphysema revealed to be the main success factors for the resolution of this complication.

This paper proposes a novel way to combine different observation models in a particle filter framework. This, so called, auto-adjustable observation model, enhance the particle filter accuracy when the tracked objects overlap without infringing a great runtime penalty to the whole tracking system. The approach has been tested under two important real world situations related to animal behavior: mice and larvae tracking. The proposal was compared to some state-of-art approaches and the results show, under the datasets tested, that a good trade-off between accuracy and runtime can be achieved using an auto-adjustable observation model.

Técnicas de agrupamento de dados usualmente assumem que o conjunto de dados é de tamanho fixo e pode ser alocado na memória. Neste contexto, um desafio consiste em aplicar técnicas de agrupamento em bases de dados de tamanho ilimitado, com dados gerados continuamente e em ambientes dinâmicos. Dados gerados nessas condições originam o que se convencionou chamar de Fluxo Contínuo de Dados (FCD). Em aplicações de FCD, operações de acesso aos dados são restritas a apenas uma leitura ou a um pequeno número de acessos aos dados, com limitações de memória e de tempo de processamento. Além disso, a distribuição dos dados gerados por essas fontes pode ser não estacionária, ou seja, podem ocorrer mudanças ao longo do tempo, denominadas de mudanças de conceito. Nesse sentido, algumas técnicas de agrupamento em FCD foram propostas na literatura. Muitas dessas técnicas são baseadas no algoritmo das k-Médias. Uma das limitações do algoritmo das k-Médias consiste na definição prévia do número de grupos. Ao se assumir que o número de grupos é desconhecido a priori e que deveria ser estimado a partir dos dados, percorrer o grande espaço de soluções possíveis (tanto em relação ao número de grupos, k, quanto em relação às partições possíveis para um determinado k) torna desafiadora a tarefa de agrupamento de dados - ainda mais sob a limitação de tempo e armazenamento imposta em aplicações de FCD. Neste contexto, essa tese tem como principais contribuições: (i) adaptar algoritmos que têm sido usados com sucesso em aplicações de Fluxo Contínuo de Dados (FCD) nas quais k é conhecido para cenários em que se deseja estimar o número de grupos; (ii) propor novos algoritmos para agrupamento que estimem k automaticamente a partir do FCD; (iii) avaliar sistematicamente, e de maneira quantitativa, os algoritmos propostos de acordo com as características específicas dos cenários de FCD. Foram desenvolvidos 14 algoritmos de agrupamento para FCD capazes de estimar o número de grupos a partir dos dados. Tais algoritmos foram avaliados em seis bases de dados artificiais e duas bases de dados reais amplamente utilizada na literatura. Os algoritmos desenvolvidos podem auxiliar em diversas áreas da Mineração em FCD. Os algoritmos evolutivos desenvolvidos mostraram a melhor relação de custo-benefício entre eficiência computacional e qualidade das partições obtidas. Several algorithms for clustering data streams based on k-Means have been proposed in the literature. However, most of them assume that the number of clusters, k, is known a priori by the user and can be kept fixed throughout the data analysis process. Besides the dificulty in choosing k, data stream clustering imposes several challenges to be dealt with, such as addressing non-stationary, unbounded data that arrives in an online fashion. In data stream applications, the dataset must be accessed in order and that can be read only once or a small number of times. In this context, the main contributions of this thesis are: (i) adapt algorithms that have been used successfully in data stream applications where k is known to be able to estimate the number of clusters from data; (ii) propose new algorithms for clustering to estimate k automatically from the data stream; (iii) evaluate the proposed algorithms according to diferent scenarios. Fourteen clustering data stream algorithms were developed which are able to estimate the number of clusters from data. They were evaluated in six artificial datasets and two real-world datasets widely used in the literature. The developed algorithms are useful for several data mining tasks. The developed evolutionary algorithms have shown the best trade-off between computational efficiency and data partition quality.

Data stream mining is an active research area that has recently emerged to discover knowledge from large amounts of continuously generated data. In this context, several data stream clustering algorithms have been proposed to perform unsupervised learning. Nevertheless, data stream clustering imposes several challenges to be addressed, such as dealing with nonstationary, unbounded data that arrive in an online fashion. The intrinsic nature of stream data requires the development of algorithms capable of performing fast and incremental processing of data objects, suitably addressing time and memory limitations. In this article, we present a survey of data stream clustering algorithms, providing a thorough discussion of the main design components of state-of-the-art algorithms. In addition, this work addresses the temporal aspects involved in data stream clustering, and presents an overview of the usually employed experimental methodologies. A number of references are provided that describe applications of data stream clustering in different domains, such as network intrusion detection, sensor networks, and stock market analysis. Information regarding software packages and data repositories are also available for helping researchers and practitioners. Finally, some important issues and open questions that can be subject of future research are discussed.