Cropland Extent Mapping in South America Global Food Security - Support Analysis m Chandra Giri, Ying Zhong January 19 th, 2016
INTRODUCTI ON
STUDY AREA AND MAPPING YEAR South America July 2009 – June 2010
METHODOLO GY
CROPLAND DEFINITION Plants harvested for food, feed, and fiber, including plantations (e.g., orchards, vineyards, coffee, tea, rubber) – GFSAD 30m CLASSIFICATION SYSTEM Cropland Fallow Non-cropland
CLASSIFICATION METHOD - randomForest Ensemble classifier Good classification performance Non-parametric classifier
RANDOMFOREST CLASSIFICATION
SEGMENTATION AND MASKING
TRAINING SAMPLE SELECTION Random selection & manual selection Tile-level selection Iterative process of training sample selection
TRAINING VARIABLES Landsat 5 and 7 L1T images, band 3, 4, and 5 acquired during the period of Nov 2009 – Feb NDVI calculated from greenest pixels composited from Landsat 5 and 7 images during the period of July 2009 – June 2010 (Maximum NDVI) Slope calculated from Shuttle Radar Topography Mission (SRTM) dataset at 1 arc-sec resolution.
POST-PROCESSING Classes reduction – cropland and non-cropland Majority filtering Manual correction
RESULTS AND DISCUSSION
PRODUCT
TRAINING VARIABLES Cropland Fallow Non- cropland Maximum NDVI
The decrease in out-of-bag (oob) error rate after randomly permuting the values of variable k of the oob samples averaged over all tress and normalized by the standard deviation VARIABLE IMPORTANCE
ACCURACY VALIDATION Tile Overall Non-cropland Cropland Fallow NA a b c NA Table 1: Out-of-bag (oob) error rate estimation of overall sample and the class of non-cropland, cropland, and fallow land cover of tile classifiers.
Limitations of the input Landsat data Cloud cover Anomaly of the Landsat 7 ETM + Scan Line Corrector ACCURACY VALIDATION Factors causing classification errors: Under-representation of landscape characteristics. Uncertainties in the training dataset. Low discrimination between land cover classes.
PRODUCT COMPARISON GlobCover 2009
PRODUCT COMPARISON GCE V1.0
MAPS COMPARISON GlobCover 2009 major cropland: ‘Post-flooding or irrigated cropland’; ‘rainfed cropland’; GlobCover 2009 major and minor cropland: ‘Post-flooding or irrigated cropland’; ‘rainfed cropland’; ‘mosaic cropland’; GCE V1.0 major cropland: ‘Irrigation major’; ‘rainfed major’; GCE V1.0 major and minor cropland: ‘Irrigation major’, ‘irrigation minor’, ‘rainfed major’, ‘rainfed minor’
ACCURACY VALIDATION Country # Total validation samples classified as cropland # Validation samples classified as non-cropland # Total validation samples Number % of total samples Number % of total samples Argentina Bolivia Brazil Chile Colombia Ecuador Paraguay Peru Uruguay Venezuela South America Table 1: Total number of validation samples and number of validation samples that are classified as cropland and non-cropland in the Cropland Extent Map in each country. Resample all maps to 500 m resolution.
Compared with GlobCover 2009 major cropland CountryProducer accuracyUser accuracy Overall accuracy Kappa statistic CroplandNon-croplandCroplandNon-cropland Argentina Bolivia Brazil Chile Colombia Ecuador Paraguay Peru Uruguay VenezuelaNA South America CountryProducer accuracyUser accuracy Overall accuracy Kappa statistic CroplandNon-croplandCroplandNon-cropland Argentina Bolivia Brazil Chile Colombia Ecuador Paraguay Peru Uruguay Venezuela South America CountryProducer accuracyUser accuracy Overall accuracy Kappa statistic CroplandNon-croplandCroplandNon-cropland Argentina Bolivia Brazil Chile Colombia Ecuador Paraguay Peru Uruguay Venezuela South America CountryProducer accuracyUser accuracy Overall accuracy Kappa statistic CroplandNon-croplandCroplandNon-cropland Argentina Bolivia Brazil Chile Colombia Ecuador Paraguay Peru Uruguay Venezuela South America Compared with GlobCover 2009 major and minor cropland Compared with GCE V1.0 major cropland Compared with GCE V1.0 major and minor cropland
CountryProducer accuracyUser accuracy Overall accuracy Kappa statistic CroplandNon-croplandCroplandNon-cropland Argentina Bolivia Brazil Chile Colombia Ecuador Paraguay Peru Uruguay Venezuela NA South America CountryProducer accuracyUser accuracy Overall accuracy Kappa statistic CroplandNon-croplandCroplandNon-cropland Argentina Bolivia Brazil Chile Colombia Ecuador Paraguay Peru Uruguay Venezuela South America GlobCover 2009 major cropland map VS GCE V1.0 major cropland map GlobCover 2009 major and minor cropland map VS GCE V1.0 major and minor cropland map
MAPS COMPARISON Reasons for poor maps agreements: Different a.cropland definition. b.mapping year c.spatial resolution
CROPLAND AREA Table 3. Cropland area (1000 ha) calculated from Cropland Extent Map. Statistics of country land area are reported by FAO. Country Cropland area (1000 ha) % of total croplandCountry land area (1000 ha) 1 Percentage (%) of cropland over total land area Argentina Bolivia Brazil Chile Colombia Ecuador French Guiana Guyana Paraguay Peru Suriname Uruguay Venezuela Total
CROPLAND AREA Country Cropland Extent Map FAO (Arable land 2 + Permanent crops 3 ) GlobCover2009 major cropland map GlobCover2009 major and minor cropland map GCE V1.0 major cropland map GCE V1.0 major and minor cropland map Argentina Bolivia Brazil Chile Colombia Ecuador French Guiana Guyana Paraguay Peru Suriname Uruguay Venezuela Total Table. The percentage (%) of cropland over total land area. 1 “ Arable land is the land under temporary agricultural crops (multiple-cropped areas are counted only once), temporary meadows for mowing or pasture, land under market and kitchen gardens and land temporarily fallow (less than five years). 2 “ Permanent crops is the land cultivated with long-term crops which do not have to be replanted for several years (such as cocoa and coffee); land under trees and shrubs producing flowers, such as roses and jasmine; and nurseries (except those for forest trees, which should be classified under "forest"). (FAOSTAT, n.d.)
CONCLUSI ON
1.30-m resolution Cropland Extent Map of South America for ,401,030 hectare cropland, 6.18% of total land area in South America % of croplands are in Brazil, and 40.05% in Argentina. 3. The ‘Maximum NDVI’ band successfully discriminates cropland and non- cropland. 4. Good classification performance of randomForest classifier. 5. Similar distribution pattern of major croplands but low statistical agreement rates with GlobCover 2009 and GCE V Google Earth Engine – powerful computation capacity.
Independent accuracy validation Irrigation or Rainfed Crop Types Intensity of Cropland FUTURE PERSPECTIVE
THANK YOU Chandra Giri Ying Zhong 01/19/2016
RANDOMFOREST CLASSIFICATION Number of decision trees to grow: 600. Variables per split: square root of the number of input variables (2). Minimum leaf population: 1. Fraction of input to bag per tree: 50% (Out-of-bag mode)