Automatic Crop Classification in Northeastern China by Improved Nonlinear Dimensionality Reduction for Satellite Image Time Series

Abstract

Accurate and timely information on the spatial distribution of crops is of great significance to precision agriculture and food security. Many cropland mapping methods using satellite image time series are based on expert knowledge to extract phenological features to identify crops. It is still a challenge to automatically obtain meaningful features from time-series data for crop classification. In this study, we developed an automated method based on satellite image time series to map the spatial distribution of three major crops including maize, rice, and soybean in northeastern China. The core method used is the nonlinear dimensionality reduction technique. However, the existing nonlinear dimensionality reduction technique cannot handle missing data, and it is not designed for subsequent classification tasks. Therefore, the nonlinear dimensionality reduction algorithm Landmark–Isometric feature mapping (L–ISOMAP) is improved. The advantage of the improved L–ISOMAP is that it does not need to reconstruct time series for missing data, and it can automatically obtain meaningful featured metrics for classification. The improved L–ISOMAP was applied to Landsat 8 full-band time-series data during the crop-growing season in the three northeastern provinces of China; then, the dimensionality reduction bands were inputted into a random forest classifier to complete a crop distribution map. The results show that the area of crops mapped is consistent with official statistics. The 2015 crop distribution map was evaluated through the collected reference dataset, and the overall classification accuracy and Kappa index were 83.68% and 0.7519, respectively. The geographical characteristics of major crops in three provinces in northeast China were analyzed. This study demonstrated that the improved L–ISOMAP method can be used to automatically extract features for crop classification. For future work, there is great potential for applying automatic mapping algorithms to other data or classification tasks.