2025
|
AIN’T - An Artificial Intelligent Network Thermometer for Measurements of Link Saturation on TCP/IP Flows Proceedings Article Marcelo R. Silva; Cesar Marcondes @inproceedings{Silva2025AintAn,
title = {AIN’T - An Artificial Intelligent Network Thermometer for Measurements of Link Saturation on TCP/IP Flows},
author = {Marcelo R. Silva and Cesar Marcondes},
url = {https://www.sige.ita.br/edicoes-anteriores/2025/TRABALHOS/Artigos-com-ISSN/AIN_T - An Artificial Intelligent Network Thermometer for measurements of link saturation on TCPIP flows.pdf},
issn = {1983-7402},
year = {2025},
date = {2025-01-01},
booktitle = {Anais do Simp\'{o}sio de Aplica\c{c}\~{o}es Operacionais em \'{A}reas de Defesa (SIGE)},
pages = {205\textendash210},
address = {S\~{a}o Jos\'{e} dos Campos, Brasil},
organization = {Instituto Tecnol\'{o}gico de Aeron\'{a}utica},
abstract = {The transmission capacity of data links is crucial for network administrators. This measure is particularly significant in operational environments where maintaining communication continuity is vital. However, the principal strategy of the most widely used tools or protocols for this purpose consists of inserting extra packets into the network and throttling its transmission capacity. Such an active strategy has the potential, even momentarily, to produce packet losses in combat support applications (SAD, for example) and crash communications on the network under analysis. Seeking to avoid network overload while measuring its saturation, this work proposes AIN’T (Artificial Intelligent Network Thermometer). AIN’T measures the level of congestion on the data link passively without inserting any data packets into the respective infrastructure. To this end, it applies MLP, LSTM, and CNN Deep Learning Networks. The results show that the models extracted from these neural network architectures can distinguish between high and low-level link saturation in an IP data network with over 99% precision.},
keywords = {Deep Learning, Passive network monitoring, Transmission Control Protocol (TCP)},
pubstate = {published},
tppubtype = {inproceedings}
}
The transmission capacity of data links is crucial for network administrators. This measure is particularly significant in operational environments where maintaining communication continuity is vital. However, the principal strategy of the most widely used tools or protocols for this purpose consists of inserting extra packets into the network and throttling its transmission capacity. Such an active strategy has the potential, even momentarily, to produce packet losses in combat support applications (SAD, for example) and crash communications on the network under analysis. Seeking to avoid network overload while measuring its saturation, this work proposes AIN’T (Artificial Intelligent Network Thermometer). AIN’T measures the level of congestion on the data link passively without inserting any data packets into the respective infrastructure. To this end, it applies MLP, LSTM, and CNN Deep Learning Networks. The results show that the models extracted from these neural network architectures can distinguish between high and low-level link saturation in an IP data network with over 99% precision. |
Automatic Detection of Airstrips in the Amazon Using ICEYE SAR Imagery and U-Net Neural Networks Proceedings Article Gustavo Henrique de Queiroz Stabile; Dimas Irion Alves; Tahisa Neitzel Kuck; Paulo Ricardo Branco Silva; Saleh Javadi @inproceedings{Stabile2025AutomaticDetection,
title = {Automatic Detection of Airstrips in the Amazon Using ICEYE SAR Imagery and U-Net Neural Networks},
author = {Gustavo Henrique de Queiroz Stabile and Dimas Irion Alves and Tahisa Neitzel Kuck and Paulo Ricardo Branco Silva and Saleh Javadi},
url = {https://www.sige.ita.br/edicoes-anteriores/2025/TRABALHOS/Artigos-com-ISSN/Automatic Detection of Airstrips in the Amazon Using ICEYE SAR Imagery and U-Net Neural Networks.pdf},
issn = {1983-7402},
year = {2025},
date = {2025-01-01},
booktitle = {Anais do Simp\'{o}sio de Aplica\c{c}\~{o}es Operacionais em \'{A}reas de Defesa (SIGE)},
pages = {164\textendash169},
address = {S\~{a}o Jos\'{e} dos Campos, Brasil},
organization = {Instituto Tecnol\'{o}gico de Aeron\'{a}utica},
abstract = {This study investigates convolutional neural networks, specifically the U-Net architecture, for the automated detection of irregular airstrips in the Amazon rainforest using Synthetic Aperture Radar (SAR) imagery from ICEYE satellites operating in the X-band. Detecting these airstrips is strategic for combating illicit activities and protecting the environment. SAR imagery is particularly effective in the Amazon, as it can penetrate cloud coverage, which is common in the region. A comprehensive pipeline was developed for data preparation, model training, and evaluation, utilizing the airstrips cataloged by the MapBiomas project for reference. Experiments were conducted by varying both the size of the input images and the balance of the training dataset. Results indicated that even with a limited number of images, the U-Net architecture can generate consistent outcomes. The study supports the development of operational solutions for the automated monitoring of irregular airstrips in the Amazon region.},
keywords = {airstrips, automatic detection, Deep Learning, Synthetic Aperture Radar},
pubstate = {published},
tppubtype = {inproceedings}
}
This study investigates convolutional neural networks, specifically the U-Net architecture, for the automated detection of irregular airstrips in the Amazon rainforest using Synthetic Aperture Radar (SAR) imagery from ICEYE satellites operating in the X-band. Detecting these airstrips is strategic for combating illicit activities and protecting the environment. SAR imagery is particularly effective in the Amazon, as it can penetrate cloud coverage, which is common in the region. A comprehensive pipeline was developed for data preparation, model training, and evaluation, utilizing the airstrips cataloged by the MapBiomas project for reference. Experiments were conducted by varying both the size of the input images and the balance of the training dataset. Results indicated that even with a limited number of images, the U-Net architecture can generate consistent outcomes. The study supports the development of operational solutions for the automated monitoring of irregular airstrips in the Amazon region. |
Classificação de áreas de garimpo ilegal na Amazônia com algoritmos de visão computacional Proceedings Article Daniel Martins Monteiro Silva; Dimas Irion Alves; Sarah Negreiros de Carvalho Leite @inproceedings{Silva2025ClassificacaoGarimpo,
title = {Classifica\c{c}\~{a}o de \'{a}reas de garimpo ilegal na Amaz\^{o}nia com algoritmos de vis\~{a}o computacional},
author = {Daniel Martins Monteiro Silva and Dimas Irion Alves and Sarah Negreiros de Carvalho Leite},
url = {https://www.sige.ita.br/edicoes-anteriores/2025/TRABALHOS/Artigos-com-ISSN/Classifica\c{c}\~{a}o de \'{a}reas de garimpo ilegal na Amaz\^{o}nia com algoritmos de vis\~{a}o computacional.pdf},
issn = {1983-7402},
year = {2025},
date = {2025-01-01},
booktitle = {Anais do Simp\'{o}sio de Aplica\c{c}\~{o}es Operacionais em \'{A}reas de Defesa (SIGE)},
pages = {182\textendash186},
address = {S\~{a}o Jos\'{e} dos Campos, Brasil},
organization = {Instituto Tecnol\'{o}gico de Aeron\'{a}utica},
abstract = {A atividade de garimpo ilegal na Amaz\^{o}nia tem crescido exponencialmente nos \'{u}ltimos anos, aumentando proporcionalmente o n\'{u}mero de \'{a}reas e a quantidade de dados que os \'{o}rg\~{a}os de controle ambiental precisam analisar. Aliado a esses fatos, o monitoramento satelital sofre com longos intervalos de revisita e condi\c{c}\~{o}es meteorol\'{o}gicas desfavor\'{a}veis, como a cobertura de nuvens persistente sobre a floresta. Para mitigar essas limita\c{c}\~{o}es, o emprego de ve\'{i}culos a\'{e}reos, como drones e aeronaves, equipados com sensores \'{o}pticos permite a aquisi\c{c}\~{a}o de imagens de alta resolu\c{c}\~{a}o. Diante do grande volume de dados, algoritmos de vis\~{a}o computacional despontam como instrumentos eficientes para classificar automaticamente essas \'{a}reas. Este estudo implementa e compara dois modelos de diferentes complexidades computacionais \textendash MobileNetV2 e EfficientNetV2-S \textendash na classifica\c{c}\~{a}o de cenas a\'{e}reas da Amaz\^{o}nia. Os resultados indicam que a MobileNetV2 apresentou o melhor desempenho com acur\'{a}cia de cerca de 98% na identifica\c{c}\~{a}o de imagens com garimpo versus sem garimpo.},
keywords = {Amaz\^{o}nia, Deep Learning, garimpo, Vis\~{a}o Computacional},
pubstate = {published},
tppubtype = {inproceedings}
}
A atividade de garimpo ilegal na Amazônia tem crescido exponencialmente nos últimos anos, aumentando proporcionalmente o número de áreas e a quantidade de dados que os órgãos de controle ambiental precisam analisar. Aliado a esses fatos, o monitoramento satelital sofre com longos intervalos de revisita e condições meteorológicas desfavoráveis, como a cobertura de nuvens persistente sobre a floresta. Para mitigar essas limitações, o emprego de veículos aéreos, como drones e aeronaves, equipados com sensores ópticos permite a aquisição de imagens de alta resolução. Diante do grande volume de dados, algoritmos de visão computacional despontam como instrumentos eficientes para classificar automaticamente essas áreas. Este estudo implementa e compara dois modelos de diferentes complexidades computacionais – MobileNetV2 e EfficientNetV2-S – na classificação de cenas aéreas da Amazônia. Os resultados indicam que a MobileNetV2 apresentou o melhor desempenho com acurácia de cerca de 98% na identificação de imagens com garimpo versus sem garimpo. |
2021
|
Classificação de Alvos em Imagens SAR com Técnicas de Machine Learning Proceedings Article Fabiano Gabriel Silva; Bruna Gregory Palm; Renato Machado @inproceedings{Fabiano2021Deepb,
title = {Classifica\c{c}\~{a}o de Alvos em Imagens SAR com T\'{e}cnicas de Machine Learning},
author = {Fabiano Gabriel Silva and Bruna Gregory Palm and Renato Machado},
url = {https://www.sige.ita.br/edicoes-anteriores/2021/st/217773_1.pdf},
year = {2021},
date = {2021-01-01},
booktitle = {Simp\'{o}sio de Aplica\c{c}\~{o}es Operacionais em \'{A}reas de Defesa 2021 (SIGE2021)},
abstract = {Atividades mar\'{i}timas que v\~{a}o desde o transporte de mercadorias at\'{e} a produ\c{c}\~{a}o de petr\'{o}leo e g\'{a}s natural s\~{a}o cada vez mais presentes em \'{a}guas brasileiras. O monitoramento dessas atividades \'{e} de fundamental import\^{a}ncia para coibir a\c{c}˜oes il\'{i}citas ou ilegais. O aumento da disponibilidade de dados de sensoriamento remoto permite que imagens de radar de abertura sint\'{e}tica (SAR - Synthetic Aperture Radar) possam ser explora- das na vigil\^{a}ncia mar\'{i}tima. Este trabalho considera o problema de classifica\c{c}\~{a}o de plataformas de petr\'{o}leo e navios localizados no litoral dos estados do Rio de Janeiro e Esp\'{i}rito Santo. Para o estudo, utilizaram-se imagens SAR com polariza\c{c}\~{a}o VH. Duas t\'{e}cnicas de machine learning foram avaliadas, a saber, Random Forest e K-nearest neighbors, com as quais pode-se obter taxas de acur\'{a}cia de 81,8% e 79,2%, respectivamente.},
keywords = {Deep Learning, Machine Learning, SAR-},
pubstate = {published},
tppubtype = {inproceedings}
}
Atividades marítimas que vão desde o transporte de mercadorias até a produção de petróleo e gás natural são cada vez mais presentes em águas brasileiras. O monitoramento dessas atividades é de fundamental importância para coibir aç˜oes ilícitas ou ilegais. O aumento da disponibilidade de dados de sensoriamento remoto permite que imagens de radar de abertura sintética (SAR - Synthetic Aperture Radar) possam ser explora- das na vigilância marítima. Este trabalho considera o problema de classificação de plataformas de petróleo e navios localizados no litoral dos estados do Rio de Janeiro e Espírito Santo. Para o estudo, utilizaram-se imagens SAR com polarização VH. Duas técnicas de machine learning foram avaliadas, a saber, Random Forest e K-nearest neighbors, com as quais pode-se obter taxas de acurácia de 81,8% e 79,2%, respectivamente. |
