Publicação
Comparison of transformer-based and convolutional neural network-based (CNN) models for remote sensing image classification
| Resumo: | Remote sensing is regarded as a rich source of data that is still very valuable for mapping (classification) and monitoring of this information using various methodologies. Convolutional Neural Networks (CNNs) are frequently employed by researchers in this field as one of the key feature extraction approaches in application to satellite pictures and have achieved good performance and efficacy. Yet, they face some problems, including the overfitting problem and the need for large datasets and expensive computational resources. As a result, it would be a good idea to experiment with various Deep Learning methods, investigate them, and compare them to existing methods while taking into account all of the elements that can affect the processing. Transformers, particularly vision Transformers, have just been offered as novel Deep Learning approaches and have demonstrated good performance in a variety of domains. Hence, experimenting with these new models would be a realistic strategy in terms of learning how they behave when processing data and what value they can contribute to the picture classification area. In relation with this, the present work aims to analyze the performance of different DL methods in classifying different land cover/land use types depending on their properties and the effect of the revolution level on this. Experimental results conducted on two main datasets “EuroSAT” and “UC Merced” indicates that some CNNs, such as, “ResNET50” and “EfficientNET B0” perform well with different resolutions while for “VGG16” and the Vision Transformer, the need for a huge amount of data for the learning task is unavoidable. |
|---|---|
| Autores principais: | Belcaid, Mareyam |
| Assunto: | CNN Transformer Vision Transformer Swin Transformer multihead attention remote sensing image classification Deep Learning |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Remote sensing is regarded as a rich source of data that is still very valuable for mapping (classification) and monitoring of this information using various methodologies. Convolutional Neural Networks (CNNs) are frequently employed by researchers in this field as one of the key feature extraction approaches in application to satellite pictures and have achieved good performance and efficacy. Yet, they face some problems, including the overfitting problem and the need for large datasets and expensive computational resources. As a result, it would be a good idea to experiment with various Deep Learning methods, investigate them, and compare them to existing methods while taking into account all of the elements that can affect the processing. Transformers, particularly vision Transformers, have just been offered as novel Deep Learning approaches and have demonstrated good performance in a variety of domains. Hence, experimenting with these new models would be a realistic strategy in terms of learning how they behave when processing data and what value they can contribute to the picture classification area. In relation with this, the present work aims to analyze the performance of different DL methods in classifying different land cover/land use types depending on their properties and the effect of the revolution level on this. Experimental results conducted on two main datasets “EuroSAT” and “UC Merced” indicates that some CNNs, such as, “ResNET50” and “EfficientNET B0” perform well with different resolutions while for “VGG16” and the Vision Transformer, the need for a huge amount of data for the learning task is unavoidable. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.