You are here:
Comparison of Sub-Pixel Classification Approaches for Crop-Specific Mapping
Citation:
Shao, Y. AND R. S. LUNETTA. Comparison of Sub-Pixel Classification Approaches for Crop-Specific Mapping . In Proceedings, 17th International Conference on Geoinformatics 2009 , Fairfax, VA, August 12 - 14, 2009. Institute of Electrical and Electronics Engineers Incorporated (IEEE), Piscataway, NJ, 1-4, (2011).
Impact/Purpose:
The Moderate Resolution Imaging Spectroradiometer (MODIS) data has been increasingly used for crop mapping and other agricultural applications. Phenology-based classification approaches using the NDVI (Normalized Difference Vegetation Index) 16-day composite (250 m) data product is among the most promising for automated processing. Most MODIS-NDVI crop mapping applications to date have focused on per-pixel classification methods; while the sub-pixel crop patterns and proportions have not been thoroughly exploited. Linear mixture model is one of the most commonly used approaches for estimating sub-pixel land-cover proportions [1], [2]. The spectral response of a pixel is assumed to be a linear combination from contributing sub-pixel land-cover proportions. The linear mixture assumption, however, might not be valid when NDVI values are considered as endmembers or inputs to linear mixture models. In addition, there might be high level iterations between MODIS-NDVI features or bands from different acquisition dates. The linear mixture model thus might not be appropriate for MODIS-NDVI sub-pixel analysis. The non-linear models such as neural network and regression tree are better suited for handling complex datasets with nonlinear relationships and high-order feature interactions [3]. Although these non-linear unmixing models have been widely used for many remote sensing applications [4], few studies have compared their performances for crop-specific sub-pixel estimations. The main objective of this paper was to implement and compare two non-linear unmixing models: (a) Multilayer Perceptron (MLP) neural network regression algorithm; and (b) Regression tree (RT) approach. The sub-pixel proportions were estimated for three major crop types including corn, soybean, and wheat; throughout the Great Lakes Basin (GLB).
Description:
This paper examined two non-linear models, Multilayer Perceptron (MLP) regression and Regression Tree (RT), for estimating sub-pixel crop proportions using time-series MODIS-NDVI data. The sub-pixel proportions were estimated for three major crop types including corn, soybean, and wheat; throughout the entire 480,000 km2 Laurentian Great Lakes Basin. Accuracy assessments were conducted using the cropland data layer (CDL) developed by the National Agricultural Statistics Service (NASS). The performances of the sub-pixel classifications were compared based on root-mean-square error (RMSE) and scatter plots. For MLP regression, the RMSE values at 500 m spatial resolution were 0.16, 0.14, and 0.07 for corn, soybean and wheat, respectively. The RT approach achieved slightly higher RMSE values of 0.18, 0.15, and 0.07 for corn, soybean, and wheat. The latter approach did not provide greater interpretability, because tree sizes were rather large for MODIS-NDVI sub-pixel crop estimation problems.
URLs/Downloads:
LUNETTA 09-052 PAPER IN NON-EPA PROCEEDING GEOINFORMATICS_PROCEEDINGS_2009.PDF (PDF, NA pp, 107 KB, about PDF)Comparison of Sub-Pixel Classification Approaches for Crop-Specific Mapping
