07 data structures_and_representations

Data structures & representations

[email protected]

mailto:[email protected]

remote sensing data processing architectures

Job Queue

Analysis Workers

Data Catalogue

Processing Workers

Auto Scaling

Ingestion

Data Catalogue

Exploitation

Annotations Catalogue

User Application Servers

LoadBalancer

User

Source Products

Domain Expert

Configuration

Admin

Domain Expert

direct data import

Data Processing / Data IntelligenceServers

13/34

Hadoop Cluster computing

[Lisa Vaas 2016]

Spark cluster computing

[Hitesh Dharmdasani, “Python and Bigdata - An Introduction to Spark (PySpark)”]

spark + mongodb

The log data structure• An append-only ordered sequence

of records.

• In DBs: log shipping protocols to transmit portions of log to slave replica databases

• In distributed systems: the State Machine Replication Principle: If two identical, deterministic processes begin in the same state and get the same inputs in the same order, they will produce the same output and end in the same state.

https://engineering.linkedin.com/distributed-systems/log-what-every-software-engineer-should-know-about-real-time-datas-unifying

Why database management systems

• An interface between the database and application programs, ensuring that data is consistently organized and remains easily accessible.

• Manages: 1. the data 2. the engine that allows data to be accessed, locked and modified 3. the database schema, which defines the database’s logical structure.

Why database management systems• Provides:

• Data abstraction and independence; Data security; A locking mechanism for concurrent access;

• An efficient handler to balance the needs of multiple applications using the same data;

• The ability to swiftly recover from crashes and errors, including restartability and recoverability;

• Robust data integrity capabilities; Logging and auditing of activity;

• Simple access using a standard application programming interface (API);

• Uniform administration procedures for data.

• Question:

• Does your application need all this? Typically yes if concurrent insert/update accesses, distributed reads…

History• 1960s: Navigational DBs

• 1970s-1980s: SQL, normalisation and OLTP, transactions

• 1990s: object oriented DBs and OLAP, warehousing

• 2000s: NoSQL and the CAP theorem

• 2010s: newSQL, graph DBs, Big Data

Databases: a user’s view• SQL and ACID: Atomicity, Consistency, Isolation, and Durability

• Variety and NoSQL: MongoDB

• Volume and NoSQL: HBase, Cassandra

• Velocity and NoSQL: MonetDB, KairosDB

• NoSQL: bad at relationships —> Graph DBs: Neo4J

• The CAP theorem: Consistency, Availability, Partition tolerance

SQL vs NoSQL

[Lisa Vaas 2016]

DB indices• Objective: sub-linear search (e.g. O(logN), O(1))

• E.g. bitmaps, keys/pointers to records, keys/pointers to blocks, reverse indices.

• Implementations in terms of (balanced) trees, hashes, B+trees.

• Types:

• non-clustered: logical order only, multiple indices possible

• clustered: physical order too for efficiency, a single clustered index per table.

KD Trees• Binary space-partitioning trees in D dimensions

• For every non-leaf node: generate a splitting hyperplane that divides the space into two half-spaces.

• Canonical construction: given all input points

• Cycles through the axes

• Split by the medians with respect to the current axissource: wikipedia

LSH

• Locality-Sensitive Hashing: maximize probability of collision —similar items end up in same bucket.

• Applications in near-duplicate detection and “fingerprinting”, similarity nearest neighbor search, hierarchical clustering.

• E.g. by random projection: use a random hyperplane to hash vectors.

Inverted indices

• Forward index: document —> content

• Inverted index: content —> document / location

• Record level, “Word” level

• Allows fast search (increases insertion cost!): queries can be resolved by jumping to the “word” id in the inverted index.

apache arrow

import feather path = 'my_data.feather' feather.write_dataframe(df, path) df = feather.read_dataframe(path)

Google Earth Engine

• A Short Intro by Kersten Clauss…

• Question: how would you replicate what’s under the hood?

18

dataacquisition&management

GeoEuskadisatelliteimageservices• linkedopendatamanagementinfrastructure• publicsources:LandsatyCopernicusSentinel1-2

19

dataprocessing

mapupdatebydistributedanalysisof25cm- regional-scaleortho-imagery

LozanoSilva,J.;AginakoBengoa,N.;Quartulli,M.;Olaizola,I.G.;Zulueta,E.,"Web-BasedSupervisedThematicMapping,"inSelectedTopicsinAppliedEarthObservationsandRemoteSensing,IEEEJournalof,vol.8,no.5,pp.2165-2176,May2015

20

dataprocessing–examplevideo

image search: iqcbm

Index management system

Data input

Column-based DB

User interface

Analysis & processing

Imageanalysis

• with DLR IMF BW• search by compression in compressed streams• dynamic taxonomies

• corel 10k

iqcbm: geoeye

TELEIOS FP7-257662

D3.1 KDD concepts and methods proposal: report & design recommendations 94

� Query-based on image content as example.

� Query-based on semantic labels.

In the case of query-based on image content, the end-user selects an image, which is passed as content example for the query. Later the system internally computes the FCD (see equation 5) between the query patch and the other patches stored in the database and display the results ordered according to the short distance. Figure 49 presents an example. In here, the user selects an image containing “houses” and the retrieved images are shown below. From the results, we can observe most of the retrieved images are correct since they contain roofs.

In the case of query-based on semantic label, the end-user introduces a semantic label and the system queries the database with this label. The result is a list of patches associated with the semantic label. An example is shown in Figure 50. Here, the user introduced “houses” as query label and the system retrieved all associated patches.

It is important to note that the patches were previously manually annotated with a semantic label and these labels are stored into the database as part of the patch information.

In the following, the results of different queries using the CBIR-FCD and TerraSAR-X images are described.

Figure 49: Results of query based on image content. Houses in Rome were selected as example and the retrieved images are presented below.

TELEIOS FP7-257662


7. Applicable�and�Reference�Documents�

7.1. Applicable�Documents��

Document Title Internal

Reference

Katrin Molch et al., (2010). “Naming Convention for Image Patches”. [RI- 1]

N. Ritter and M. Ruth, (1995) “GeoTIFF Format Specification GeoTIFF”, Revision 1.0, Version 1.8.1. [RI- 2]

Robert M Haralick, K Shanmugam, Its'hak Dinstein (1973). "Textural Features for Image Classification". IEEE Transactions on Systems, Man, and Cybernetics, SMC-3 (6): 610–621.

[RI- 3]

The GLCM Tutorial http://www.fp.ucalgary.ca/mhallbey/tutorial.htm [RI-4]

H.G. Feichtinger, Th. Strohmer (1998). "Gabor Analysis and Algorithms", Birkhäuser. [RI-5]

B.S. Manjunath, W.Y. Ma (1996). “Texture features for browsing and retrieval of image data”. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18 (8): 837–842.

[RI-6]

A. Popescu, I. Gavat, M. Datcu (2008). "Complex SAR image characterization using space variant spectral analysis". IEEE Radar Conference, 1-4. [RI-7]

T. Li, M. Ogihara (2006). “Towards Intelligent Music Information Retrieval”. IEEE Transactions on Multimedia, 8(3), 564-574. [RI-8]

A. Croisier, D. Esteban, C. Galand (1976). "Perfect channel splitting by use of interpolation / decimation / tree decomposition techniques". International Conf. on information Science and Systems, Patras, Greece: 443-446.

[RI-9]

P.P Vaidyananthan (1987). “Quadrature Mirror Filter Banks, M-Band Extensions and Perfect-Reconstruction Techniques”. IEEE ASSP Magazine, 4 (3): 4-20.

[RI-10]

M. Campedel, E. Moulines, M. Datcu (2005). “Feature Selection for Satellite Image Indexing”. ESA-EUSC: Image Information Mining – Theory and Application to Earth Observation.

[RI-11]

N. Karthikeyani Visalakshi, K. Thangavel (2009). “Impact of Normalization in Distributed K-Means Clustering”. International Journal of Soft Computing, 4 (4):168-172.

[RI-12]

D. Dai, W. Yang, H. Sun (2011). “Multilevel Local Pattern Histogram for SAR Image Classification,” IEEE Geosciences Remote Sensing Letters, 8 (2): 225–229.

[RI-13]

S. Cui, M. Datcu, (2011) “Coarse to Fine Patches-based Multitemporal Analysis of Very High Resolution Satellite Images”, Multitemp 2011, to be published. [RI-14]

D. Cerra, M. Datcu, (2010). “Image Retrieval using Compression-based Techniques”, Proceedings of the International ITG Conference on Source and Channel Coding, Siegen, Germany.

[RI-15]

Welch, T. (1984). “Technique for high-performance data compression”, Computer 17(6): 8-19. [RI-16]

23

iqcbm: digitalglobe TELEIOS FP7-257662





Reference




[RI- 3]




[RI-6]




[RI-9]


[RI-10]


[RI-11]


[RI-12]


[RI-13]



[RI-15]


24

.76

.69

.79 .90

.95

.99

TELEIOS FP7-257662





Reference




[RI- 3]




[RI-6]




[RI-9]


[RI-10]


[RI-11]


[RI-12]


[RI-13]



[RI-15]


normalized distance from query

25

iqcbm: terrasar-x TELEIOS FP7-257662





Reference




[RI- 3]




[RI-6]




[RI-9]


[RI-10]


[RI-11]


[RI-12]


[RI-13]



[RI-15]


26

iqcbm: terrasar-x TELEIOS FP7-257662





Reference




[RI- 3]




[RI-6]




[RI-9]


[RI-10]


[RI-11]


[RI-12]


[RI-13]



[RI-15]


27

experiment: tsx

iqcbm: terrasar-x

TELEIOS FP7-257662


Table 10: Test data set description. Both scenes, Venice and Toulouse were tiled into patches with 200 pixel sizes.

TSX MGD-SE product

Total patches

Patch size Semantic labels

Venice 1054 200x200 18

Toulouse 1170 200x200 17

The patches were annotated with a semantic label by using the Search Engine based on SVM tool and user supervision previously presented in section 5.1. The semantic labels associated to the selected classes were previously described in Table 5.

In the following, we present some examples of retrieving TerraSAR-X structures using both images. Table 11 displays the query images and the 20 top retrieved images. Some quality metrics (Precision and Recall) were computed from these results and they are summarized in Table 12.

Query images

Retrieved images

Class9

Class6

Class7

Class36

TELEIOS FP7-257662


Class20

Class31

Class28

Class32

Table 11: Results of the queries based on image content using CBIR-FCD as data mining tool.

Table 12 shows the precision and recall for the classes and the query time in seconds needed for searching and retrieving the results.

Table 12: Precision and recall of the semantic classes using query based on content and the query time.

Class Precision (%)

Recall (%)

Query time (sec)

Class1 5,36 5,17 0.32882 Class2 10,71 10,34 0.318238 Class3 5,36 5,17 0.235323 Class4 7,14 6,90 0.107209 Class5 8,93 8,62 0.371344 Class6 0,00 0,00 0.258542 Class7 1,79 1,72 0.206243 Class8 8,93 8,62 0.383762 Class9 41,07 39,66 0.230424 Class10 0,00 0,00 0.236996

TELEIOS FP7-257662





Reference




[RI- 3]




[RI-6]




[RI-9]


[RI-10]


[RI-11]


[RI-12]


[RI-13]



[RI-15]


TELEIOS KDD concepts and methods proposal: report & design recommendations

Corneliu Octavian Dumitru, Daniela Espinoza Molina, Shiyong Cui, Jagmal Singh, Marco Quartulli, Mihai Datcu 2011, FP7 TELEIOS Tech Report

29

interactiveweb-baseddataretrieval

web-basedinteractiveclassificationofimagecontent

07 data structures_and_representations

Science

Transcript of 07 data structures_and_representations