Cropland Extraction Using Water Mask and Phenological Features based on Landsat-8 Imagery

Abstract

Accurate cropland information is crucial for food security assessment. At present, Landsat-8 remote sensing imagery and Google Earth Engine (GEE) cloud platform provide favorable conditions for mapping cropland. Despite the improved mapping conditions, the existence of mixed pixels is still one of the main difficulties in the cropland extraction. Especially in paddy fields, the influence of water bodies on farmland extraction should not be ignored. Also, the lack of time series data with temporal-spatial resolution impedes the accurate cultivated land mapping. Aiming at these problems, we proposed an effective method to extract cropland precisely based on the Landsat-8 imagery and GEE platform. First, we created cloud-free images and masked out water based on Landsat-8 images. Second, we constructed the normalized difference vegetation index time series data with 30-m resolution from Landsat-8 and MODIS images to obtain accurate phenological information. Third, other cropland features such as texture features, tasseled cap transformation features, terrain features and spectral features were acquired. Finally, we applied the Random Forest algorithm and obtained the cropland for 2015 in Qinzhou. The results show that terrain characteristics and phenological characteristics have the greatest influence on cropland extraction. Based on water mask, the overall accuracy of cropland improved by 6.9%. Overall, the product has high accuracy with the mean overall accuracy (OA) of 97.9%. Compared with MCD12Q1 and GFSAD30 cropland products, it is 6.2% and 24.3% higher in OA respectively. In brief, our method is helpful for the cropland extraction.