B43A-0536
Mapping Crop Patterns in Central US Agricultural Systems from 2000 to 2014 Based on Landsat Data: To What Degree Does Fusing MODIS Data Improve Classification Accuracies?

Thursday, 17 December 2015
Poster Hall (Moscone South)
Likai Zhu1, Volker Radeloff1, Anthony R Ives2 and Brandon Barton3, (1)University of Wisconsin Madison, Department of Forest and Wildlife Ecology, Madison, WI, United States, (2)University of Wisconsin Madison, Department of Zoology, Madison, WI, United States, (3)Mississippi State University, Department of Biological Sciences, Mississippi State, MS, United States
Abstract:
Deriving crop pattern with high accuracy is of great importance for characterizing landscape diversity, which affects the resilience of food webs in agricultural systems in the face of climatic and land cover changes. Landsat sensors were originally designed to monitor agricultural areas, and both radiometric and spatial resolution are optimized for monitoring large agricultural fields. Unfortunately, few clear Landsat images per year are available, which has limited the use of Landsat for making crop classification, and this situation is worse in cloudy areas of the Earth. Meanwhile, the MODerate Resolution Imaging Spectroradiometer (MODIS) data has better temporal resolution but cannot capture fine spatial heterogeneity of agricultural systems. Our question was to what extent fusing imagery from both sensors could improve crop classifications. We utilized the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) algorithm to simulate Landsat-like images at MODIS temporal resolution. Based on Random Forests (RF) classifier, we tested whether and by what degree crop maps from 2000 to 2014 of the Arlington Agricultural Research Station (Wisconsin, USA) were improved by integrating available clear Landsat images each year with synthetic images. We predicted that the degree to which classification accuracy can be improved by incorporating synthetic imagery depends on the number and acquisition time of clear Landsat images. Moreover, multi-season data are essential for mapping crop types by capturing their phenological dynamics, and STARFM-simulated images can be used to compensate for missing Landsat observations. Our study is helpful for eliminating the limits of the use of Landsat data in mapping crop patterns, and can provide a benchmark of accuracy when choosing STARFM-simulated images to make crop classification at broader scales.