关键词:
Transfer learning
Random Forest
Cropland Data Layer (CDL)
USA
Cotton
Corn
TIME-SERIES DATA
RANDOM FOREST
FEATURE-SELECTION
FOOD-DEMAND
NDVI DATA
LANDSAT
AREA
PERFORMANCE
ACCURACY
UNCERTAINTY
摘要:
Training samples is fundamental for crop mapping from remotely sensed images, but difficult to acquire in many regions through ground survey, causing significant challenge for crop mapping in these regions. In this paper, a transfer learning (TL) workflow is proposed to use the classification model trained in contiguous U.S.A. (CONUS) to identify crop types in other regions. The workflow is based on fact that same crop growing in different regions of world has similar temporal growth pattern. This study selected high confidence pixels across CONUS in the Cropland Data Layer (CDL) and corresponding 30-m 15-day composited NDVI time series generated from harmonized Landat-8 and Sentinel-2 (HLS) data as training samples, trained the Random Forest (RF) classification models and then applied the models to identify crop types in three test regions, namely Hengshui in China (HS), Alberta in Canada (AB), and Nebraska in USA (NE). NDVI time series with different length were used to identify crops, the effect of time-series length on classification accuracies were then evaluated. Furthermore, local training samples in the three test regions were collected and used to identify crops (LO) for comparison. Results showed that overall classification accuracies in HS, AB and NE were 97.79%, 86.45% and 94.86%, respectively, when using TL with NDVI time series of the entire growing season for classification. However, LO could achieve higher classification accuracies earlier than TL. Because the training samples were collected across USA containing multiple growth conditions, it increased the potential that the crop growth environment in test regions could be similar to those of the training samples; but also led to situation that different crops had similar NDVI time series, which caused lower TL classification accuracy in HS at early-season. Generally, this study provides new options for crop classification in regions of training samples shortage.