Defining Remote Sensing Remote Sensing is a technology for sampling electromagnetic radiation to acquire and interpret non-immediate geospatial data from which to extract information about features, objects, and classes on the Earth's land surface, oceans, and atmosphere. (NASA)
Introduction to Remote Sensing Spatial, spectral, temporal resolutions Image display alternatives Vegetation Indices Image classifications Image change detections Accuracy assessment
Selected RS Systems Landsat TM & Enhanced TM (Thematic Mapper) Ikonos & Quickbird MODIS (Moderate Resolution Imaging Spectrometer) Radarsat & ERS (European RS Satellite) SeaWiFS (Sea-Viewing Wide Field-of-View Sensor) GOES (Geostationary Operational Environ. Satellite) AVHRR (Advanced Very High Resolution Radiometer) SPOT (High Resolution Visible)
Digital Satellite Data Remote sensing: the acquisition of data about an object or area by a sensor that is physically far from the object or area. Digital satellite data is one type of remotely-sensed data aerial photography is another type of remotely-sensed geographic information Important advantage of remote sensing data: synoptic perspective: comprehensive view of large areas of the Earth s surface, vantage point of observation, digital format, multispectral, multi-temporal, multi-spatial.
Satellite Data - Benefits & Problems Benefits: it is already digital (unlike standard aerial photos) -- readily lends itself to manipulation and analysis to extract useful information synoptic perspective repeat coverage of areas minimal scale and geometric distortion problems relative to aerial photos Problems: generally low spatial resolution compared to aerial photos large data storage volumes distortions in data values due variations in atmospheric conditions, clouds, terrain, solar angle, etc... -- require preprocessing to correct!
Electromagnetic Spectrum (EMS) Visible Portion At a given Temperature, T 0.4 0.5 0.6 0.7 Cosmic Gamma X-rays UV Infrared Microwave Radio/TV Rays Rays Near/Medium/Thermal 10-6 10-4 10-2 1 10 2 10 4 10 6 10 8 Wavelength in Micrometers (= 10 3 nanometers) Increasing Wavelength Increasing Frequency and Energy
Passive (Optical) Sensor Technology Incoming EMR Reflected EMR Emitted EMR
Electromagnetic Radiation Spectrum Remote sensing satellites record data on electromagnetic radiation (EMR) satellites have detectors that record specific wavelengths in the EM spectrum EM radiation interacts with physical matter some wavelengths are absorbed & others reflected determine/estimate matter type by analyzing spectral signatures in satellite data
The Four RS Resolutions Satellite data characterized by four resolutions: spatial resolution: area on ground represented by each pixel (cell) in the satellite data. temporal resolution: how frequently a satellite platform obtains imagery of a particular area, orbital characteristics. spectral resolution: specific EM spectrum wavelength intervals captured/recorded by a particular satellite sensor. radiometric resolution: number of possible data values recorded/reportable by each satellite sensor, precision with which the EM spectral values are reported.
Satellite Data - Pixel Brightness Values Brightness Value (BV) corresponding to intensity of EM radiation in specified spectral range detected for that pixel lower BV = lower level of EM radiation detected higher BV = higher level of EM radiation detected Displaying/viewing satellite data: high BV medium BV low BV For displaying/viewing satellite images: high BVs assigned bright/light color value low BVs assigned dark/dim color value
Satellite Data - Pixel Brightness Values & Display Landsat TM Multi-Spectral Display : RED GUN GREEN GUN BLUE GUN Band 4 - near infrared Band 3 - red visible Band 2 - green visible on-screen display: + + = Band 7 - middle infrared Band 4 - near infrared Band 2 - green visible on-screen display: + + =
RGB Compositing Process Infrared band (4) Red color gun Red band (3) Green color gun Green band (2) Blue color gun
Multispectral Composite Display Near infrared (red gun), red (green gun), green (blue gun): false color
Multispectral Composite Display Middle infrared (red gun), near infrared (green gun), green (blue gun): false color
Different Spatial Resolutions 1-2 m 30 m 79 m 1.1 km QuickBird, Landsat TM, Landsat MSS AVHRR IKONOS ETM
Digital Image Data MSS 1973 TM 1999 IKONOS 1999
Image Ratioing Image ratioing for change detection is based on the assumption that the ratio of the DN values for a stable feature over two dates should be unity (or 1), while changed pixels should have a ratio significantly different from unity. R ij = DN DN ij ij ( t ( t Due to various external factors, the stable features may not have an expected value of unity from the ratio. A critical step is to determine the threshold value to differentiate change/nochange, which is often empirical. No from-to information on change detection is available. 2 1 ) )
Vegetation Indices Normalized Difference Vegetation Index (NDVI): NDVI = ρ ρ NIR NIR ρ + ρ Commonly used index; most extensively used for global vegetation monitoring using AVHRR & MODIS data. Usually a two week or monthly composite AVHRR or MODIS data are used for vegetation monitoring. Day 1 Day n NDVI Find maximum NDVI red red (weekly, biweekly or monthly) Composite NDVI
Soil Adjusted Vegetation Index (SAVI) SAVI = (1 + L) ρ NIR ρred ρ + ρ + NIR red L Where L is an adjustment factor for soil; Huete (1988) found the optimal value for L is 0.5.
Green Fractional Coverage (fc)
Classifications Supervised Unsupervised Hybrid Classify Historical Imagery in a Time- Series Pixel vs. Object Classifications
Computer-Aided Classification Parametric: assumes a normal distribution of the data. (1) Supervised Classification: provides the computer with some examples of known features in multi-dimensional feature space. The computer will first analyze the statistical parameters for the training data and then assign all other pixels to one of the classes based on statistical similarity. (2) Unsupervised Classification: the user allows the computer to derive pre-specified numbers of spectral clusters in which the differences among clusters are maximized and the differences within clusters are minimized. It is the users responsibility to assign a class label to each of the clusters. One class may have many clusters.
Supervised Classification Algorithms 1. Minimum-distance-to-means classifier The training set provides a mean value for each class for each band, the mean vector. Each pixel is assigned to a class to which it has the shorted Euclidean distance. Band y u u u u u u u s s s s u u u s s s s u u u u w w w w w w w w f f c c c c f f f f f c c Band x
Unsupervised Classification Unsupervised classification does not use training data as the basis for classification, but examines unknown pixels in an image and aggregates them into a number of clusters based on natural groupings present in the image values. The classification assumes that DN values within a given cover type should be close together in the spectral space, while data in different classes should be comparatively well separated. The categories that unsupervised classification identify are spectral classes or clusters. The analyst must provide labels to each of the clusters after unsupervised classification using other sources of data, including air-photos and direct field observation and a GPS.
K-Means Approach iterations Iterations of recalculating the means causes the mean points to migrate to new locations of cluster center.
Different Temporal Resolutions Seasonal differences Annual differences Decadal differences Change detections & animations Rates and trajectories of landscape change Change in spatial patterns over time-space
General Change Detection Approaches Channel/Scene Integration: near-infrared channels from scenes of different periods composited as a qualitative method of assessing regional change. Multi-date Composite: multi-scene data stack representing different time periods used as the input or feature set for an unsupervised classification for defining change and no-change spectral clusters. Image Algebra: two channels of the same spectral region and wavelengths for two different time periods are ratioed, and image differencing is achieved by subtracting the spectral responses of one date from that of the other. Binary Mask: uses a multi-date image composite recoded into a binary mask consisting of areas, which have changed and not changed between two dates.
Change Detection Techniques Univariate Image Differencing The most straight forward way to see whether a change has happened is to take a difference between two images collected over the same place at different times. D ij = DN ij ( t2) DNij ( t1) The variable to use can be NDVI for vegetation change detection, or a single band reflectance. Difference is computed on a pixel basis. If geometric correction between the two images is poor, spurious change may be identified. Theoretically, if D deviates from zero, a change has happened. There may be a systematic shift in the threshold value indicating change or a stable pixels caused by sensor degradation, atmospheric conditions, phenology, and more. Determining the threshold for change and non-change is the critical step for this technique.
More Change Detection Approaches Post-Classification: classification of two scenes from different dates assessed on a pixel-by-pixel basis and reported through a change matrix -- two scenes at a time; result is a change image for each date pair, containing a number of change pairs equal to the square of the number of classes minus some number of illogical change pairs (e.g. Water-Primary Forest). Panel Data Analysis: stacks all classification dates together to generate a pixel life history. Hypothesis is that the nature of the trajectory is associated with the function of the land at that pixel and its neighborhood of similarly related pixels.
Multi-Date Composite Analysis This method links multi-date images as if they were a single date image. Image of date 1 n bands Image of date 2 n bands Composite image 2n bands
Post - Classification Change Straight-forward approach that compares classified maps for 2-time periods on a pixel by pixel basis to identify changes. Date 1 class map Date 2 class map Change detection map
Pixel Trajectories, II
Accuracy Assessment Need to evaluate the classes with independent data. Depending upon the data you have or time and resources available, you may evaluate you map using the following data approaches: 1. Existing map with known accuracy. 2. Air photos of know specifications. 3. Map in the field and compare with the real world. The data collected in the field with regard to its category are called ground truth or control data. Of the three methods, the last one is the most reliable, but it is the most costly and time consuming approach.
Accuracy Assessment: Ground Step 1. Sampling Truth Usually there are millions of pixels in a selected study area. It is impossible to check each of the pixels in the field. Rather, we only check a small representative portion of the pixels on the ground. To maintain objectivity, we choose pixels through some sort of random sampling. Simple random sampling from all the pixels may miss some important small classes. Therefore, we often do a stratified random sampling. Taking each class as a stratum and drawing a simple random sample from each class. Therefore, you know that each class is appropriately sampled; sometimes stratify by access or some other constraint.
Classification Error Matrix Step 2. The classification error matrix is also called the confusion matrix or contingency table. Here is an example error matrix: Reference data W S F U C H row total Classified W 226 0 0 12 0 1 239 S 0 216 0 92 1 0 309 F 3 0 360 228 3 5 599 U 2 108 2 397 8 4 521 C 1 4 48 132 190 78 453 H 1 0 19 84 36 219 359 Column total 233 328 429 945 238 307 2480 W:water, S: sand, F: forest, U: urban, C: corn, H: hay
Map Accuracies Producer s Accuracy: the percentage of pixels in the reference data for a certain class that are correctly identified by the classifier. User s Accuracy: the percentage of pixels classified as a certain class that agrees with the reference data. Overall Accuracy: the percentage of the total pixels that are correctly identified. Producer s Accuracy W=226/233=97% S=216/328=66% F=360/429=84% U=397/945=42% C=190/238=80% H=219/307=71% User s Accuracy W=226/239=94% S=216/309=70% F=360/599=60% U=397/521=76% C=190/453=42% H=219/359=61% Overall Accuracy=(226+216+360+397+190+219)/2480=65%
K_hat Statistics If we close our eyes and randomly assign the pixels to the classes, we still would have some of them put in the correct class. K_hat statistic is a measure of difference between the map accuracy the accuracy and random assignment. ˆ = k Observed accuracy - chance agreement 1- chance agreement K_hat is a numerical measure of the extent to which the percentage correct values of an error matrix are due to true agreement versus chance agreement. K_hat varies between 0 and 1. 0 means the classification is no better than random classification, and 1 means true agreement. For example, K_hat=0.67 means that the classification is 67% better than randomly assign the pixels to the classes.
Accuracy Assessment of Land- Change Models Treat each observation (e.g., a pixel) as something meaningful to the phenomena. Assume each observation is independent in space and time. Statistical approaches need to address landscape persistence and no-change pixels. Precision of data (e.g., composition & pattern): spatial resolution, magnitude, location.
Why Sub-Pixel Classification or Spectral Unmixing? Relationship between the sensor IFOV (instantaneous field of view) & size of object. The boundaries between objects. Land use/land cover proportions or fractions in a remote sensing cell resolution. Spectral mixing is always present; the necessity of sub-pixel classification is largely dependent on the scene characteristic and the users interests.
Active Remote Sensing Rather than rely upon incident solar radiation, a signal is sent toward the target landscape May be satellite-borne or airborne Includes some RADAR systems as well as LIDAR
What is LIDAR? LIght Detection And Ranging (laser altimeter) Capable of generating high resolution Digital Elevation Models (DEMs) Horizontal resolution dependent upon pulse rate, aircraft altitude, and aircraft velocity Vertical resolution is centimeter level with a RMS of 15 cm** **(Instrument specifications)
Recent Research Objectives Apply detailed LIDAR surveys to a variety of research problems including: Coastal erosion studies Geomorphology / slope stability Land cover classification & Vegetation structure Watershed studies Archaeology Improve analysis techniques such as: Vegetation removal and analysis Feature extraction and incorporation of intensity data Improve accuracy and data quality through: Developing survey procedures for specific types of surveys Improving GPS accuracy
OPTECH 1225 AIRBORNE LASER TERRAIN MAPPER (ALTM) LIDAR INSTRUMENT Elevation accuracy 15 cm at 3,600 ft AGL Laser footprint variable from 10 to 20 cm depending on aircraft elevation Laser pulse rate: up to 25,000 per sec. Laser scan rate: variable up to 50 Hz. Laser scan angle: variable up to ±20 Operating altitude: 1,000-6,000 ft AGL Swath width: up to 4,000 ft at 6,000ft AGL. Installable in a variety of aircraft including a single engine Cessna 206. Laser scan lines ALTM laser and IMU GPS satellites Aircraft GPS GPS ground reference station Flight direction
Why LIDAR? Comparison of DEMs A) Standard USGS 30m 30m digital elevation model (DEM) B) AverStar 10m 10m DEM created from USGS 7.5 minute / 1:24,000 topos C) Vegetation-filtered 0.5m 0.5m DEM derived from LIDAR data of Austin (Vegetation removed Buildings remain)