Reference Data Standard Protocols Kamini Yadav Dr. Russ Congalton.

Slides:



Advertisements
Similar presentations
E-Agri Project Kick-off Meeting, Mol, 24-25, 2011 Remote Sensing of Crop Acreage and Crop Mapping in the E-Agri Project Chen Zhongxin Institute of Agricultural.
Advertisements

Months of the year December January November October February
How Information on a Map Can Be Displayed
VEGETATION MAPPING FOR LANDFIRE National Implementation.
Multi-temporal and Multi-resolution Analysis of Normalized Differential Vegetation Index and Rainfall towards Global Irrigated Area Mapping 1. Introduction.
North American Croplands Richard Massey & Dr. Teki Sankey.
Multiple Criteria for Evaluating Land Cover Classification Algorithms Summary of a paper by R.S. DeFries and Jonathan Cheung-Wai Chan April, 2000 Remote.
VALIDATION OF REMOTE SENSING CLASSIFICATIONS: a case of Balans classification Markus Törmä.
The Role of RS Techniques in European Land Use Database Construction Centre for Geo-Information 1 The Role of RS Techniques in European Land.
Self-Supervised Segmentation of River Scenes Supreeth Achar *, Bharath Sankaran ‡, Stephen Nuske *, Sebastian Scherer *, Sanjiv Singh * * ‡
Lecture 14: Classification Thursday 18 February 2010 Reading: Ch – 7.19 Last lecture: Spectral Mixture Analysis.
P Pathophysiology Calendar. SundayMondayTuesdayWednesdayThursdayFridaySaturday January 2012.
Accuracy Assessment and Reference data Collection Kamini Yadav Dr. Russ Congalton.
Overview Minimum required classifiers for mapping vegetation cover at global scale using the FAO-LCCS tool GLC LEGEND hjs/30-Apr-01.
James C. Tilton Code Computational & Information Sciences and Technology Office NASA Goddard Space Flight Center January 30, 2015 update National.
Accuracy Assessment. 2 Because it is not practical to test every pixel in the classification image, a representative sample of reference points in the.
Common Core Where have we been and where we are going…
North American Croplands: US Updates 21 May 2015 Richard Massey; Dr. Teki Sankey.
F.A.O. Land Cover/Remote Sensing Specialist
LMD/IPSL 1 Ahmedabad Megha-Tropique Meeting October 2005 Combination of MSG and TRMM for precipitation estimation over Africa (AMMA project experience)
Phase I Forest Area Estimation Using Landsat TM and Iterative Guided Spectral Class Rejection Randolph H. Wynne, Jared P. Wayman, Christine Blinn Virginia.
National Mapping Division EROS Data Center U. S. Geological Survey U.S. Geological Survey Earth Resources Operation Systems (EROS) Data Center World Data.
U.S. Department of the Interior U.S. Geological Survey Web Presence, Data Sharing, Real- time Analysis and Crowdsourcing GFSAD30 Sixth Workshop – July.
Lu Liang, Peng Gong Department of Environmental Science, Policy and Management, University of California, Berkeley And Center for Earth System Science,
Scenes of the Earth: Slide Show Assessment Activity.
Realities of Satellite Interpretation (The things that will drive you crazy!) Rachel M.K. Headley, PhD USGS Landsat Project.
Selection of Multi-Temporal Scenes for the Mississippi Cropland Data Layer, 2004 Rick Mueller Research and Development Division National Agricultural Statistics.
U.S. Department of the Interior U.S. Geological Survey GFSAD30 Field Work Planning: Progress in Australia Pardha, Prasad, and Jun GFSAD30 monthly meeting,
WORD JUMBLE. Months of the year Word in jumbled form e r r f b u y a Word in jumbled form e r r f b u y a february Click for the answer Next Question.
Map of the Great Divide Basin, Wyoming, created using a neural network and used to find likely fossil beds See:
Remotely sensed land cover heterogeneity
BMTRY 789 Lecture 11: Debugging Readings – Chapter 10 (3 rd Ed) from “The Little SAS Book” Lab Problems – None Homework Due – None Final Project Presentations.
Reference Data Standard Protocols Kamini Yadav Dr. Russ Congalton.
14 ARM Science Team Meeting, Albuquerque, NM, March 21-26, 2004 Canada Centre for Remote Sensing - Centre canadien de télédétection Geomatics Canada Natural.
DATE POWER 2 INCOME JANUARY 100member X 25.00P2, FEBRUARY 200member X 25.00P5, MARCH 400member X 25.00P10, APRIL 800member.
Assessing the Phenological Suitability of Global Landsat Data Sets for Forest Change Analysis The Global Land Cover Facility What does.
1 Psych 5500/6500 Measures of Variability Fall, 2008.
Accuracy Assessment: Building Global Cropland Reference Data Updates for March 2015 Kamini Yadav and Russ Congalton.
U.S. Department of the Interior U.S. Geological Survey GFSAD30 Field Work Planning: Progress in Australia Pardha, Prasad, and Jun GFSAD30 monthly meeting,
U.S. Department of the Interior U.S. Geological Survey Monthly Progress for Africa GCEV2 Jun Xiong, Prasad, Pardha 23 October, 2014.
BOT / GEOG / GEOL 4111 / Field data collection Visiting and characterizing representative sites Used for classification (training data), information.
Cropland using Google Earth Engine
1 Global Mapping, Approaches, Issues, and Accuracies Russ Congalton & Kamini Yadav University of New Hampshire January 16, 2014 Menlo Park, CA.
U.S. Geological Survey U.S. Department of Interior GFSAD 30 Cropland Products of Nominal 250 m Using MODIS Data and Cropland Mapping Algorithms:
2011 Calendar Important Dates/Events/Homework. SunSatFriThursWedTuesMon January
Realities of Satellite Interpretation (The things that will drive you crazy!) Rachel M.K. Headley, PhD USGS Landsat Project.
Object-Based Crop Identification for Collecting Reference Data from VHRI RHSeg and Ecognition Software Kamini Yadav and Russ Congalton.
Road Trip Across the United States. How Great Are Sports!?!
Optical Properties in coastal waters change rapidly on very fine spatial scales. The existence of multiple ocean color systems provides a unique capability.
Mapping Canada’s Rangeland and Forage Resources using Earth Observation Emily Lindsay MSc Candidate – Carleton University Supervisors: Doug J. King & Andrew.
Reference data & Accuracy Assessment Dr. Russ Congalton Kamini Yadav.
Dr. Russ Congalton & Kamini Yadav GFSAD30 Meeting, Menlo Park 19 th -21 st January, 2016 Reference Data Collection & Accuracy Assessment: Some Results.
U.S. Geological Survey U.S. Department of Interior Automated Cropland Mapping Algorithms (ACMA) for Global Food Security Assessment Across Years Pardhasaradhi.
Website: Croplands.org. User Authentication Full user authentication system with roles. User passwords hashed (one way encrypted) with a salt using modern.
US Croplands Richard Massey Dr Teki Sankey. Objectives 1.Classify annual cropland extent, Rainfed-Irrigated, and crop types for the US at 250m resolution.
ACCURACY ASSESSMENT SOUTH AMERICA, NORTH AMERICA KAMINI YADAV DR. RUSSELL CONGALTON.
26. Classification Accuracy Assessment
Temporal Classification and Change Detection
An Ecosystem Services Approach to Water Resources
Assessment of Current Field Plots and LiDAR ‘Virtual’ Plots as Guides to Classification Procedures for Multitemporal Analysis of Historic and Current Landsat.
Map of the Great Divide Basin, Wyoming, created using a neural network and used to find likely fossil beds See:
Evaluating Land-Use Classification Methodology Using Landsat Imagery
Office of Education Improvement and Innovation
Hands on Satellite Data to Monitor Biomass
Summary Probability of a cloud free image no more than 16 and 32 days apart during the growing season Still using a fixed growing season April 1st to Oct.
Supervised Classification
McDonald’s calendar 2007.
McDonald’s calendar 2007.
2015 January February March April May June July August September
Presentation transcript:

Reference Data Standard Protocols Kamini Yadav Dr. Russ Congalton

Current Process Flow Chart ???this probably still need some tweak, after final discussions???

Testing Protocol on Mali data Evaluate Mali data collected in August 2015 by Murali, according to the flowchart made by Justin Large Scale, LS (>20ha) Medium Scale, MS(10-20 ha) Small Scale, SS(<10ha) Very Large Scale, VLS Very Very Large Scale, VVLS Field collected Created by Justin ???we need to define scale taking pixel resolution of satellite sensors: LS >=6.25 ha (MODIS 250m) but = 0.81 ha (3x3, 30 m) =100 ha (AVHRR 1000 m) but = 1600 ha???

Mali Ground Data (August 2015, Murali) Need to know/understand…… How to handle multiple crops in single sample? ??crop dominance (e.g., wheat 60%, barley 20%, others 20%) rather than crop type??? What does scale mean here? ??in previous slide we defined scale with their relationship to satellite sensor pixel sizes; so whether we map crop types or crop dominance, scale definitions are same?? How does this scale help in deciding homogeneous samples at 250m/90m sample unit? ??pure homogeneous samples for crop type are not feasible across the world especially at resolutions >30 m (0.09 ha). So, where we cant get pure single crop homogenity, we should define by crop dominance (e.g., wheat-barley system; corn-soybean system)?? What the date of the collection of crop data? ??as far as possible data is collected based on crop phenology. Again this maynot be always possible??

Steps to crosswalk the Mali data into our standard protocol format Extract the single crop samples because they conform to the homogeneity requirement due to the presence of a single crop ??In Africa, it is unlikely to find single crop even within a Landsat pixel (30 m x 30 m or 0.09 ha). So, where crop type is not feasible crop dominance is the way to go. Remember we have 4 products. For product 1 (crop no crop) it is possible to get sufficient number of homogeneous pixels, same for product 2 (irrigated vs. rainfed), where homogeneous crops are mapped we can get crop calendar for that (product 3). However, product 4 we may not be able to map crop type everywhere, but crop dominance)?? Check the homogeneity of each sampling unit in a 250mX 250m window on Google Earth (using high resolution imagery) ??homonenety should mean getting croplands right. Not crop type. Because I expect most pixels to have mixed crops at 250 m resolution in Africa. However, this is possible in USA, Canada, Europe??

Checking Mali data using Google Earth The images of some samples do not look like a single crop ??this will be most common for 250 m pixel. You will have crop dominance or mix of crops of various proportions?? Some samples might not have crops before the growing season (left as fallow) ??such data is useful for identifying crop, no crop?? Are the coordinates correct (are we in the right place)? ??likely coordinates are taken from road. People who use the data have to correct for this by moving the pixels to field centers??

Issue: Proximity to Road 2014 Need to implement some automated filter on the entire dataset based on some minimum distance to the road layer ?? People who use the data have to correct for this by moving the pixels to field centers….applying automated filter is even better (Justing, please let us know)?? All the samples near to road (within certain distance) must be flagged for evaluation or moved to the center of a field ??indeed. See previous response….also everyone I know have been doing this??

April 2014 November 2013 Issue - continued ??These are common problems of field work. People who use the data have to correct for this by moving the pixels to field centers??

Issue - continued February 2014 VVLS?? – very, very large scale Does this mean only homogeneous crops are there so that we can move the sample inside the field and away from the road? ??VVLS defined in slide 3. The above filed is homogeneous for MODIS 250m) but = 0.81 ha (3x3, 30 m) <=6.25 ha); SS = <=0.81 ha (3x3, 30 m)???:

2014 February 2015 VLS -Very Large scale?? We need to understand this scale term so that we can determine how far the samples can be moved away from the road and into the field and yet remain a homogeneous sample unit. ??see definition of scale in slide 3 and answers previously?? Issue - continued

March2013 Includes settlement in the yellow window. This is Large scale (LS) but does not look homogeneous??? ??Here is an example on how to deal with these. When one gets reference data, let us say thet have pure homogeneous pixels for class 1 : croplands, single crop, rainfed, LS (sample size N=22). The non pure samples are not used in classification, but used just in class identification. For example, the above pixel inspite of having a partial pixel in settlement, still may classify into class 1 : croplands, single crop, rainfed, LS. So, it helps in class identification and labeling?? Issue - Close to settlement

February 2011 June 2013 Include settlements in the yellow window ??samples such as this will be useful in class labeling. Suppose one of your unsupervised (e.g., class 56) Happens to be this class, then we can label the class As: croplands mixed with settlements??? Question - So close to settlements may be indication of the wrong geographic coordinates?? ?

September 2013 January 2014 The sample includes a waterbody in 250m by 250m homogeneous window? ??samples such as this will be useful in class labeling. Suppose one of your unsupervised (e.g., class 56) Happens to be this class, then we can label the class As: croplands mixed with water bodies or it may actually fall into : croplands, single crop, irrigated, rice dominant???

December 2013 Not homogeneous ??will be still valuable to identify classes?? Close to the coastline

Photograph taken of the sample unit does not look like Maize crop as mentioned in description ??not sure…ask murali?? February 2014 December2011

February 2014 Homogeneity vs. Scale Large scale?? In Google Earth, the 250x250m window does not look homogeneous; Photograph seems to be of more than one crops ??answered previously??

VLS…does this scene indicate homogeneity for one crop type?? ??probably it has one or more crop with about 20% tree cover?? Issue - homogeneity

Important Issues to be solved Identify the single crop sample units ??this should be attempted. But, landscape is diverse and fragmented. Probably if you drive in Mali, getting single crop fields of modis size (250m x 250m) will be very rare. It is possible to get 3 x 3, 30 m pixels (0.81 ha) as homogeneous unit. But, if the ladscape is all mixed crop, we should capture it as is?? Decide homogeneous samples from different scale levels – very confusing ??homogenity is possible for: (a) crop vs. no crop, (b) irrigated vs. rainfed. For crop type, homogenity may or maynot exist. All our field work is standardized for 250m x 250 m pixels (6.25 ha). In Africa, only a very small proportion of pixels will be homogeneous at this resolution. The same may well be true for even a 3 x 3 30m Landsat pixels (0.81 ha). In such cases we should capture mixed crops, just as is the reality in field (we can’t change this)?? Perform stratification and identify samples for each class label ?? If Justin can do this, terrific. However, if there are 1000 points collected, we split samples by 60 (600):40 (400) and share them with reference and validation teams. The people who do reference data work, validation data work can do this stratification themselves as well. But, all ground data points are precious and need to be retained. They may or maynot be homogeneous for crop type, but that is reality in field conditions?? In addition to minimum distance between samples (spatial autocorrelation), use minimum distance to roads, settlements, and waterbodies. This analysis must be added to the flow chart. ??ok. Something to keep in mind that when people gather data, they travel a minimum of 2-3 kilometers from one point to another…..so, this issue does not remove samples??

Thank you Questions?