Improving the mapping of crop types in the Midwestern U.S. by fusing Landsat and MODIS satellite data

Mapping crop types is of great importance for assessing agricultural production, land-use patterns, and the environmental effects of agriculture. Indeed, both radiometric and spatial resolution of Landsat’s sensors images are optimized for cropland monitoring. However, accurate mapping of crop types requires frequent cloud-free images during the growing season, which are often not available, and this raises the question of whether Landsat data can be combined with data from other satellites. Here, our goal is to evaluate to what degree fusing Landsat with MODIS Nadir Bidirectional Reflectance Distribution Function (BRDF)-Adjusted Reflectance (NBAR) data can improve crop-type classification. Choosing either one or two images from all cloud-free Landsat observations available for the Arlington Agricultural Research Station area in Wisconsin from 2010 to 2014, we generated 87 combinations of images, and used each combination as input into the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) algorithm to predict Landsat-like images at the nominal dates of each 8-day MODIS NBAR product. Both the original Landsat and STARFM-predicted images were then classified with a support vector machine (SVM), and we compared the classification errors of three scenarios: 1) classifying the one or two original Landsat images of each combination only, 2) classifying the one or two original Landsat images plus all STARFM-predicted images, and 3) classifying the one or two original Landsat images together with STARFM-predicted images for key dates. Our results indicated that using two Landsat images as the input of STARFM did not significantly improve the STARFM predictions compared to using only one, and predictions using Landsat images between July and August as input were most accurate. Including all STARFM-predicted images together with the Landsat images significantly increased average classification error by 4% points (from 21% to 25%) compared to using only Landsat images. However, incorporating only STARFM-predicted images for key dates decreased average classification error by 2% points (from 21% to 19%) compared to using only Landsat images. In particular, if only a single Landsat image was available, adding STARFM predictions for key dates significantly decreased the average classification error by 4 percentage points from 30% to 26% (p < 0.05). We conclude that adding STARFM-predicted images can be effective for improving crop-type classification when only limited Landsat observations are available, but carefully selecting images from a full set of STARFM predictions is crucial. We developed an approach to identify the optimal subsets of all STARFM predictions, which gives an alternative method of feature selection for future research.