A Terrain Segmentation System for SAR Images Using Compact Convolutional Neural Networks
Ahishali, Muharrem Mete (2019)
2019
Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2019-05-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201905211692
https://urn.fi/URN:NBN:fi:tty-201905211692
Tiivistelmä
Accurate terrain segmentation of synthetic aperture radar (SAR) images plays an important role in environmental, economic and natural research areas and applications. For example, terrain segmentation can be used for oil spill detection, vegetation analysis, and land cover/land usage mapping. There are several methods proposed in this domain consisting of traditional Machine Learning (ML) methods and deep Convolutional Neural Networks (CNNs).
The traditional ML methods generally focus on designating highly discriminative features to improve the segmentation performance. The hand-crafted features are derived from Target Decomposition theorems (TDs) over SAR data and its second order descriptors such as the coherency and covariance matrices. Moreover, as the recent trend, image processing features: color and texture have started to be used together with electromagnetic (EM) features derived from TDs. Nonetheless, the integration and utilizing such high-dimensional features for terrain segmentation are not straightforward due to the well-known phenomena “the curse of dimensionality” and causes degradations on the segmentation performance.
On the other hand, deep CNNs are promising the state-of-the-art (soa) performance levels in many computer vision tasks such as image recognition, segmentation, and tracking without requiring future extraction process. However, the studies using deep CNNs have certain major limitations and drawbacks such as the requirement of a huge amount of training data as in “Big Data” scale and special hardware requirement for the implementation of computationally complex architectures.
In this thesis work, a novel and systematic terrain segmentation framework is proposed based on a compact and adaptive implementation of CNNs using sliding-window classification approach. The proposed approach can overcome the abovementioned limitations of deep CNNs whilst achieving the soa segmentation performance levels. Overall, the novel and innovative contributions of this work can be listed as the following. First, the proposed approach can perform the segmentation task by using directly backscattering coefficients or the covariance and coherency matrices; hence, there is no need for additional feature extraction process unlike the conventional ML methods which usually utilize a large set of features. Second, the proposed compact CNNs are very efficient in terms of computational complexity because of their compact nature compared to its deep counterparts. This efficiency is also valid against the conventional methods, since there is no requirement for additional feature extraction process in the proposed approach. Next, the proposed approach achieves the soa segmentation performance levels with the minimum human-expertise interaction and costly labelling (e.g., usually requires labeled data <0.1% of the SAR data). This is substantially important in SAR terrain segmentation since the available labeled data is scarce in practice. Therefore, the necessity of an automated terrain segmentation, unfortunately, becomes questionable, if three-quarter of the data is labelled which is the case in many existing deep CNN type of approaches in remote sensing. Finally, the proposed approach is able to perform segmentation using small window sizes which is unfeasible with deep CNNs. Hence, it requires no post-processing such as filtering for blocking artifact.
The experimental evaluations are performed over six benchmark SAR data including both airborne and spaceborne sensors: four fully polarimetric, one single- and one dual-polarized data. Hence, the experimental results over these benchmark datasets show how the proposed approach can overcome the aforementioned drawbacks and limitations including the limitations due to single- or dual-polarimetry. The quantitative and qualitative extensive set of performance comparisons reveal that the proposed approach outperforms both the conventional ML and deep CNN based approaches with minimum computational complexity and human labelling.
The traditional ML methods generally focus on designating highly discriminative features to improve the segmentation performance. The hand-crafted features are derived from Target Decomposition theorems (TDs) over SAR data and its second order descriptors such as the coherency and covariance matrices. Moreover, as the recent trend, image processing features: color and texture have started to be used together with electromagnetic (EM) features derived from TDs. Nonetheless, the integration and utilizing such high-dimensional features for terrain segmentation are not straightforward due to the well-known phenomena “the curse of dimensionality” and causes degradations on the segmentation performance.
On the other hand, deep CNNs are promising the state-of-the-art (soa) performance levels in many computer vision tasks such as image recognition, segmentation, and tracking without requiring future extraction process. However, the studies using deep CNNs have certain major limitations and drawbacks such as the requirement of a huge amount of training data as in “Big Data” scale and special hardware requirement for the implementation of computationally complex architectures.
In this thesis work, a novel and systematic terrain segmentation framework is proposed based on a compact and adaptive implementation of CNNs using sliding-window classification approach. The proposed approach can overcome the abovementioned limitations of deep CNNs whilst achieving the soa segmentation performance levels. Overall, the novel and innovative contributions of this work can be listed as the following. First, the proposed approach can perform the segmentation task by using directly backscattering coefficients or the covariance and coherency matrices; hence, there is no need for additional feature extraction process unlike the conventional ML methods which usually utilize a large set of features. Second, the proposed compact CNNs are very efficient in terms of computational complexity because of their compact nature compared to its deep counterparts. This efficiency is also valid against the conventional methods, since there is no requirement for additional feature extraction process in the proposed approach. Next, the proposed approach achieves the soa segmentation performance levels with the minimum human-expertise interaction and costly labelling (e.g., usually requires labeled data <0.1% of the SAR data). This is substantially important in SAR terrain segmentation since the available labeled data is scarce in practice. Therefore, the necessity of an automated terrain segmentation, unfortunately, becomes questionable, if three-quarter of the data is labelled which is the case in many existing deep CNN type of approaches in remote sensing. Finally, the proposed approach is able to perform segmentation using small window sizes which is unfeasible with deep CNNs. Hence, it requires no post-processing such as filtering for blocking artifact.
The experimental evaluations are performed over six benchmark SAR data including both airborne and spaceborne sensors: four fully polarimetric, one single- and one dual-polarized data. Hence, the experimental results over these benchmark datasets show how the proposed approach can overcome the aforementioned drawbacks and limitations including the limitations due to single- or dual-polarimetry. The quantitative and qualitative extensive set of performance comparisons reveal that the proposed approach outperforms both the conventional ML and deep CNN based approaches with minimum computational complexity and human labelling.