Download presentation
Presentation is loading. Please wait.
Published byEthelbert Todd Modified over 8 years ago
1
Mapping Canada’s Rangeland and Forage Resources using Earth Observation Emily Lindsay MSc Candidate – Carleton University Supervisors: Doug J. King & Andrew M. Davidson
2
Research Objectives To explore potential methods and data sources useful for distinguishing between forage land cover types, and to use this knowledge to produce a cost-effective and accurate methodology for distinguishing land cover types within Canada’s rangeland and forage resources using remotely sensed imagery and geo- spatial data. Random Forest – Implementation of advanced pixel-based supervised classification method – Analysis of variable importance – Comparing Out-of-Bag error to independent validation Optical Data: Vegetation Indices + Phenological Variables – Testing the use of additional variables derived from optical imagery to enhance overall accuracy SAR Data – Test if RADARSAT-2 imagery improves classification accuracy when used alongside Landsat-8 imagery Acquisition Timing – Testing the required number and timing of optical and radar imagery Training Data – Comparing ground data sources Provincial Crop insurance datasets Field surveys
3
Annual Space-Based Crop Inventory for Canada Decision Tree (DT) methodology Landsat-8 Radarsat-2 Overall target accuracy of 85% Satellite Images Ground Data Classification and Processing Validation Annual Crop Inventory Accuracy of the Annual Crop Map
4
Study Sites
5
Data Collection GPS Enabled ArcPad Tablet
6
Class Description
7
Imagery Data EO Data Sources Study SiteSensorDate of Acquisition ManitobaLandsat 8 OLIMay 27, 2015 August 30, 2015 October 18, 2015 RADARSAT-2 (Wide)July 8 – July 15, 2015 AlbertaLandsat 8 OLIMay 19, 2015 July 6, 2015 August 23, 2015 RADARSAT-2 (Wide)July 20 – 27, 2015
8
Data: Phenology and Vegetation Index Variables
9
VI Example: TCW Spring ImageMid Summer Image Seeded Forage
10
Phenology Example: NDVI StDev
11
Random Forest Classification Optical + Radar Variables Random Forest Classifier 2/3 Test Set 1/3 Test Set Accuracy Assessment Reduction of Variables Classified Map
12
Results: Alberta Variables: MS (3), VI (3), Phenology + Radar Overall Accuracy94.00 Cropland PA97.97 Rangeland PA91.67 Seeded Forage PA88.18 Cropland UA96.02 Rangeland UA96.35 Seeded Forage UA86.61 OOB Error93.73 Kappa0.90
13
Results: Manitoba Variables: MS (3) + VI (3) + Phenology + Radar Overall Accuracy 88.48 Cropland PA 98.48 Rangeland PA 60.47 Seeded Forage PA 86.99 Cropland UA 97.37 Rangeland UA 80.00 Seeded Forage UA 77.44 OOB Error 90.66 Kappa 0.81
14
Classifier Performance 36 Total Classifications Best Single Date Alberta MS + VI + Radar (Late Summer) – 90% Overall Accuracy Manitoba MS + VI + Radar (Spring) – 87% Overall Accuracy Best 2 Date Alberta MS + VI + Radar (Spring & Late Summer) – 93% Overall Accuracy Manitoba MS + VI + Radar (Spring & Fall) – 89% Overall Accuracy Best 3 Date Alberta VI + Phenology + Radar – 94.6% Overall Accuracy Manitoba MS + Radar – 89.5% Overall Accuracy
15
Future Work – Operational Prairie Rangeland and Forage Dataset Alberta Saskatchewan
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.