Machine Learning For Detecting Anomalies In SAR Data

  • Slides: 16
Download presentation
Machine Learning For Detecting Anomalies In SAR Data Yuval Haitman Itay Berkovich Supervisor: Prof.

Machine Learning For Detecting Anomalies In SAR Data Yuval Haitman Itay Berkovich Supervisor: Prof. Stanley Rotman 1

Outline • Introduction • SAR temporal data • Algorithm implementation • Results and validation

Outline • Introduction • SAR temporal data • Algorithm implementation • Results and validation • Conclusion • Questions 2

Introduction • Research Domain: anomaly detection in remote sensing multi-dimensional imagery • What will

Introduction • Research Domain: anomaly detection in remote sensing multi-dimensional imagery • What will be defined as an anomaly? • Difficulties of anomaly detectors: • • Choosing a proper decision threshold • Dealing with noise Our goal: develop an algorithm for anomaly detection that will use data adaptive thresholds and will be less affected by noise 3

SAR Data 4

SAR Data 4

RX Anomaly Detector • Common anomaly detector for multi-dimensional data • Assumption: • Image

RX Anomaly Detector • Common anomaly detector for multi-dimensional data • Assumption: • Image background follows multivariate normal distribution • Variations: • Global RX (GRX) • Local RX (LRX) 5

Non-Negative Matrix Factorization (NNMF) Sparse NNMF: 6

Non-Negative Matrix Factorization (NNMF) Sparse NNMF: 6

Non-Negative Matrix Factorization (NNMF) • The extracted features are trends in data • Each

Non-Negative Matrix Factorization (NNMF) • The extracted features are trends in data • Each endmembers has unique time signature • The data is spanned using those trends 7

Non-Negative Matrix Factorization (NNMF) Original Data Reduced Data 8

Non-Negative Matrix Factorization (NNMF) Original Data Reduced Data 8

Algorithm Overview 9

Algorithm Overview 9

Preprocessing 10

Preprocessing 10

Dimension Reduction and RX Detector 12

Dimension Reduction and RX Detector 12

Thresholding method 13

Thresholding method 13

Results and Validation • We used Google Earth for result validation • Figure (a)

Results and Validation • We used Google Earth for result validation • Figure (a) – SAR image with marked anomalies • Figure (b)- marked clusters on Google Earth image (a) (b) 14

Palmachim Set – Result Validation Shafdan pools (Red cluster) 17/10/11 29/01/12 Yes Planet –Rishon

Palmachim Set – Result Validation Shafdan pools (Red cluster) 17/10/11 29/01/12 Yes Planet –Rishon lesion (Yellow cluster) 11/02/10 30/03/11 17/10/11

Results and Validation – Extended cluster 16

Results and Validation – Extended cluster 16

Conclusions • We have implemented a new algorithm that is able to detect anomalies

Conclusions • We have implemented a new algorithm that is able to detect anomalies which other algorithms could not • The algorithm should be tested on more data sets in order to explore it’s reliability and sensitivity of detection 18