Uncertainty analysis of rice planting area extraction based on different classifiers using Landsat data

Abstract: Rice is the staple food for over half of the world's population and two-thirds of the population of China. One of the main methods to implement an estimate of the planting area is to classify an image of the study area. Systematic quality assessment and some quantitative researches have been made on uncertainties in rice area estimation using remote sensing data. In this paper, sub-meter GPS data from a field campaign and TM image of study area were combined to obtain 1m resolution sub-pixels of simulated images. Maximum Likelihood Classifier(MLC), K-Nearest Neighbors (KNN), BP neural network (BPN) and Fuzzy ARTMAP neural network (FUZZY ARTMAP) were used as hard classification approaches to classify the TM image of the study area. Classification results showed that the classification precision of all non-parametric approaches (KNN,BPN and FUZZY ARTMAP) were higher than that of parametric approach (MLC). The differences of overall accuracy between these three non-parameters classifications were small. As for rice area, it's better to choose MLC to get higher User's Accuracy, and choose KNN to get higher Producer's Accuracy. Full fuzzy BPN, partial fuzzy BPN and KNN classifiers were used to estimate area of classes in sub-pixels of simulated and TM images. The accuracies of area estimation by full fuzzy BPN classifier were significantly higher than these by partial fuzzy BPN and KNN classifiers. The correlation coefficient between the predicted area and true area of sub-pixels was not suitable in accuracy assessment for fuzzy classification, but a paired t-test could be used to assess well accuracy of area estimation. Full fuzzy classifiers have advantages of selecting eligible and enough training samples over partial fuzzy classifiers and enhance classification precision. But classification results failed to offer different categories of each pixel in space in the location information. The combined multiple classifiers either in voting mode or in measuring mode showed capacities to enhance the overall classification uncertainty in this study. It can help to improve the precision of the rice area extraction to some extent. An approach to analyzing the mixing degree of pixels was proposed in this study. The mixing degree of pixels of 30m resolution TM image was calculated by up scaling thematic map on majority rule in Matlab. As far as the condition of rice growing regions in southern China is concerned, the problem of mixed pixel is much more severe for commonly used images like TM images. And the classification results demonstrated that the classification precision decreased with the pureness of pixels and four classifiers showed no difference in capacity to classify mixing pixels. Based on Probability Vector which was available to BPN and KNN classifiers, the maps of maximum probability, entropy of all pixels and probability of pixels with rice label were made to represent uncertainties of classification for the TM image of the study area. These maps with the traditional classification map can transfer not only results of classification but also information of spatial variation of classification uncertainty to users.