TY - JOUR
T1 - Mapping abandoned agriculture with multi-temporal MODIS satellite data
AU - Alcantara, Camilo
AU - Kuemmerle, Tobias
AU - Prishchepov, Alexander
AU - Radeloff, Volker C.
PY - 2012
Y1 - 2012
N2 - Agriculture is expanding and intensifying in many areas of the world, but abandoned agriculture is also becoming more widespread. Unfortunately, data and methods to monitor abandoned agriculture accurately over large areas are lacking. Remote sensing methods may be able to fill this gap though, especially with the frequent observations provided by coarser-resolution sensors and new classification techniques. Past efforts to map abandoned agriculture relied mainly on Landsat data, making it hard to map large regions, and precluding the use of phenology information to identify abandoned agriculture. Our objective here was to test methods to map abandoned agriculture at broad scales with coarse-resolution satellite imagery and phenology data. We classified abandoned agriculture for one Moderate Resolution Imaging Spectroradiometer (MODIS) tile in Eastern Europe (~1,236,000km 2) where abandoned agriculture was widespread. Input data included Normalized Difference Vegetation Index (NDVI) and reflectance bands (NASA Global MODIS Terra and Aqua 16-Day Vegetation Indices for the years 2003 through 2008, ~250-m resolution), as well as phenology metrics calculated with TIMESAT. The data were classified with Support Vector Machines (SVM). Training data were derived from several Landsat classifications of agricultural abandonment in the study area. A validation was conducted based on independently collected data. Our results showed that it is possible to map abandoned agriculture for large areas from MODIS data with an overall classification accuracy of 65%. Abandoned agriculture was widespread in our study area (15.1% of the total area, compared to 29.6% agriculture). We found strong differences in the MODIS data quality for different years, with data from 2005 resulting in the highest classification accuracy for the abandoned agriculture class (42.8% producer's accuracy). Classifications of MODIS NDVI data were almost as accurate as classifications based on a combination of both red and near-infrared reflectance data. MODIS NDVI data only from the growing-season resulted in similar classification accuracy as data for the full year. Using multiple years of MODIS data did not increase classification accuracy. Six phenology metrics derived with TIMESAT from the NDVI time series (2003-2008) alone were insufficient to detect abandoned agriculture, but phenology metrics improved classification accuracies when used in conjunction with NDVI time series by more than 8% over the use of NDVI data alone. The approach that we identified here is promising and suggests that it is possible to map abandoned agriculture at broad scales, which is relevant to gain a better understanding of this important land use change process.
AB - Agriculture is expanding and intensifying in many areas of the world, but abandoned agriculture is also becoming more widespread. Unfortunately, data and methods to monitor abandoned agriculture accurately over large areas are lacking. Remote sensing methods may be able to fill this gap though, especially with the frequent observations provided by coarser-resolution sensors and new classification techniques. Past efforts to map abandoned agriculture relied mainly on Landsat data, making it hard to map large regions, and precluding the use of phenology information to identify abandoned agriculture. Our objective here was to test methods to map abandoned agriculture at broad scales with coarse-resolution satellite imagery and phenology data. We classified abandoned agriculture for one Moderate Resolution Imaging Spectroradiometer (MODIS) tile in Eastern Europe (~1,236,000km 2) where abandoned agriculture was widespread. Input data included Normalized Difference Vegetation Index (NDVI) and reflectance bands (NASA Global MODIS Terra and Aqua 16-Day Vegetation Indices for the years 2003 through 2008, ~250-m resolution), as well as phenology metrics calculated with TIMESAT. The data were classified with Support Vector Machines (SVM). Training data were derived from several Landsat classifications of agricultural abandonment in the study area. A validation was conducted based on independently collected data. Our results showed that it is possible to map abandoned agriculture for large areas from MODIS data with an overall classification accuracy of 65%. Abandoned agriculture was widespread in our study area (15.1% of the total area, compared to 29.6% agriculture). We found strong differences in the MODIS data quality for different years, with data from 2005 resulting in the highest classification accuracy for the abandoned agriculture class (42.8% producer's accuracy). Classifications of MODIS NDVI data were almost as accurate as classifications based on a combination of both red and near-infrared reflectance data. MODIS NDVI data only from the growing-season resulted in similar classification accuracy as data for the full year. Using multiple years of MODIS data did not increase classification accuracy. Six phenology metrics derived with TIMESAT from the NDVI time series (2003-2008) alone were insufficient to detect abandoned agriculture, but phenology metrics improved classification accuracies when used in conjunction with NDVI time series by more than 8% over the use of NDVI data alone. The approach that we identified here is promising and suggests that it is possible to map abandoned agriculture at broad scales, which is relevant to gain a better understanding of this important land use change process.
KW - Agricultural abandonment
KW - Change detection
KW - Eastern Europe and the former Soviet Union
KW - Fallow land
KW - Farmland
KW - Land use and land cover change
KW - Landsat
KW - MODIS
KW - Phenology
KW - Support vector machines
KW - SVM
KW - Time series
U2 - 10.1016/j.rse.2012.05.019
DO - 10.1016/j.rse.2012.05.019
M3 - Journal article
AN - SCOPUS:84862242893
SN - 0034-4257
VL - 124
SP - 334
EP - 347
JO - Remote Sensing of Environment
JF - Remote Sensing of Environment
ER -
