Track 2: Water detection and classification on multi-source remote sensing and terrain data

Introduction

This processing track addresses a particular problem in the field of big data processing and management with the objective of simulating a realistic remote sensing application scenario. The main focus is on the detection of water surfaces (natural waters, flood, inland excess water, other water-affected categories) using remotely sensed data. Multiple independent data sources are available and different tools could be used for data processing and evaluation. The main challenge is to identify the right combination of data and methods to solve the problem in the most efficient way.

The first task is to determine the most useful input source and the processing methodology most useful for the problem. In real-life situations a careful optimization is needed in order to not only obtain the best possible achievements in terms of processing accuracy, but also from the resource and time efficiency viewpoints. In generic terms, one has to evaluate the available inputs and processing methodologies and to select an optimum solution that can fit into the available time frame with the available resources (computing, human).

In this track a challenge is defined to address the problem of detecting water surfaces and different categories of water-affected soils and vegetation, covered by a variety of remotely sensed data of multiple resolutions.

IQPC 2015 will be the theme of a special session in the ISPRS GeoBigData conference (http://www.isprs-geospatialweek2015.org/workshops/geobigdata/) and track reports will be reviewed to be included in the conference proceedings.

Background

Detection and monitoring of various water surfaces has been a challenge for a long time in remote sensing data processing. A large number of studies are available in the relevant literature dealing with water and wetness detection and monitoring for land management and conservation. Numerous different types of RS data are useful for some kind of water detection, however the accuracy is highly dependent on the input source, the processing methodology and in particular the combination of the two.

In real-life situations, a balance is achieved by creating a processing chain consisting of different methods and input sources and spatially aware algorithms are used to combine the tools to provide a result of sufficient information content. However a very challenging task is to optimize the resources for the tasks.

Objectives

The goal of this track is to detect water-related thematic classes in a specific area in Hungary.

The track leader provides a wide array of remote sensing data covering the area in concern, including:

  • high-resolution aerial hyperspectral imagery,
  • high-resolution visible (RGB) orthophotos,
  • Terrain model (DTM) and surface model (DSM) derived from airborne LiDAR point clouds,
  • medium-resolution (Landsat 8) satellite imagery

The participants have to provide thematic maps with a set of pre-defined categories. A set of calibration and validation samples will be provided to train and validate the various processing chains, and final evaluation will be carried out by the track leader using independent samples.

Competitive solutions have to fulfil the below criteria:

  • Create the best possible classification using the simplest set/combination of input sources
  • Try to reduce the number of input data for the processing
  • Develop algorithms that are fast to run
  • Find the best balance of complexity and accuracy (maximize efficiency) during the data processing
Practical implementation

The study area is divided into three parts:

  • Area A: the full study area with medium-resolution Landsat 8 coverage
  • Areas B and C: two sub-areas with full high-resolution data coverage.

The participants have to provide thematic maps for each of the areas (A, B and C) with the below categories:

  • Water surfaces (code: 1)
  • Wet/waterlogged soils (code: 2)
  • Soils not directly affected by water (code: 3)
  • Vegetation standing in water (code: 4)
  • Vegetation not directly affected by water (code: 5)
  • Other (code: 0)

For each area, a set of training and verification samples will be provided by the track leader in vector format, covering each of the above thematic categories. The participants can use the samples to train the classification algorithms and to verify the results. Areas with both high- and low-resolution data coverage (B and C) can be used to tune the algorithms for better performance on the larger area with only Landsat data (area A).

Participants have to submit the following material:

  • Georeferenced thematic rasters in GeoTIFF format, containing the codes of thematic categories as described above
  • Concise description of the whole methodology and processing chain (including algorithms and parameters, with references to relevant literature wherever available)
Data description

Location: The input datasets cover a study area in Hungary. The area of interest is on the North-East part of the country along the river Bodrog. The lowest point of the area is on 77, 19 m above the sea level, the highest point is on 258,73 m above the sea level.

 

Location of the datasets

Landsat 8 data is provided for the whole area A, whereas hyperspectral images, LiDAR-derived Digital Surface (DSM) and Terrain Model (DTM) and orthophotos (RGB) are additionally provided for areas B and C.

 

Hyperspectral images

Hyperspectral images are provided as georeferenced (UTM34N) radiance data in .dat (ENVI) format containing 128 bands.

Spectral and spatial resolution and accuracy:

Instrumentation / camera: aerial hyperspectral instrument (AISA Eagle)
Spectral range: approx. 400-1000 nm (visible and reflected (near) infrared, VIS / VNIR)
Number of channels: 128
Spectral resolution: 5 nm
Spatial resolution: 1.5 m / pixel
Spatial accuracy: 2.5 m (RMSE)
Data content: spectral radiance
Data Type / Format: Georeferenced  ENVI DAT file; 16 bit BSQ
Projection: UTM/WGS84, zone 34 North

 

Digital Terrain Data (DTM)

The Digital Terrain Model is generated from aerial laser scanner (ALS) data, with gap filling by appropriate interpolation.

Spatial resolution and accuracy:

Spatial resolution: 1 meter / pixel
Spatial accuracy: 30 cm
Data content: height values in cm, interpreted over the Baltic basic sea level.
Data Type / Format: raster (unsigned long integer) (.tif)
Projection: UTM/WGS84, zone 34 North

Digital orthophotos

Spectral and spatial resolution and accuracy:
Spectral range: Visible (RGB)
Number of Bands: 3
Spatial resolution: 15 cm / pixel
Spatial accuracy of 30 cm (RMSE)
Data Type / Format: Radiometrically and geometrically corrected images in TIFF
Projection: UTM/WGS84, zone 34 North

Landsat 8 Data

Landsat 8 data products are consistent with the all standard Level-1 (orthorectified) data products with the following specifications:

Processing: Level 1 T- Terrain Corrected

Pixel Size: OLI multispectral bands 1-7,9: 30-meters, TIRS thermal bands 10-11 acquired at 100 m and resampled to 30 meters

OLI panchromatic band 8: 15-meters

Data Characteristics of the original dataset:

GeoTIFF data format, Cubic Convolution (CC) resampling, North Up (MAP) orientation, Universal Transverse Mercator (UTM) map projection, World Geodetic System (WGS) 84 datum, 12 meter circular error, 90% confidence global accuracy for OLI, 41 meter circular error, 90% confidence global accuracy for TIRS, 16-bit pixel values

Band description: Landsat 8 Operational Land Imager (OLI) images consist of nine spectral bands with a spatial resolution of 30 meters for Bands 1 to 7 and 9. New band 1 (ultra-blue) is useful for coastal and aerosol studies. New band 9 is useful for cirrus cloud detection. The resolution for Band 8 (panchromatic) is 15 meters. Thermal Infrared Sensor (TIRS) bands 10 and 11 are useful in providing more accurate surface temperatures and are collected at 100 meters.

 

 

 

 

 

Landsat 8
Operational
Land Imager
(OLI)
and
Thermal
Infrared
Sensor
(TIRS)

Launched
February 11, 2013

Bands

Wavelength
(micrometers)

Resolution
(meters)

Band 1 - Coastal aerosol

0.43 - 0.45

30

Band 2 - Blue

0.45 - 0.51

30

Band 3 - Green

0.53 - 0.59

30

Band 4 - Red

0.64 - 0.67

30

Band 5 - Near Infrared (NIR)

0.85 - 0.88

30

Band 6 - SWIR 1

1.57 - 1.65

30

Band 7 - SWIR 2

2.11 - 2.29

30

Band 8 - Panchromatic

0.50 - 0.68

15

Band 9 - Cirrus

1.36 - 1.38

30

Band 10 - Thermal Infrared (TIRS) 1

10.60 - 11.19

100 ( resampled to 30)

Band 11 - Thermal Infrared (TIRS) 2

11.50 - 12.51

100 (resampled to 30)

Details about processing levels of all Landsat data products can be found here.

For the area of interest we provide a subset of a Landsat 8 imagery what need to be used for the processing.

Data download

Data can be downloaded from here: http://map.fomi.hu/download/IQPC_15_T2/

Evaluation

The evaluation and scoring will be developed based upon the complexity, time- and resource efficiency of the methodology and the data requirement for processing. This means that the best result will be the simplest solution that is fast and easy to carry out, requires less input, however it is accurate enough to address the problem.

The track leader will assess the results submitted by each participant by using a verification sample set independent of those provided to the participants.

Detailed evaluation protocol will be provided soon.

Timeline

[April  1, 2015]: Announcement of the IQmulus processing contest @ website

[June 26, 2015]: Deadline for submissions

[July 1, 2015]: Contest organizers report the evaluations

[August  3, 2015]: Announcement of the contest award

[Sept 28, 2015]: Presentation of the contest at the workshop