Exploring intra-annual variation in cropland classification accuracy using monthly, seasonal, and yearly sample set

Abstract

Monitoring cropland dynamics is critical for water management and crop-yield estimation but cropland mapping is still a challenge because of its seasonality and human intervention. Both large-area mapping and change detection studies use images from certain or the same seasons; however, the uneven distribution of Landsat data within a year may affect accuracy. This study aimed to explore the accuracy and uncertainty of cropland mapping performance in different season and month by analyzing all available images in three Landsat footprints in Egypt, Ethiopia, and South Africa from 1984 to 2016. Three different combinations of training models (using yearly, seasonal, and monthly sample set) were used to compare the effect on cropland classification in different month and season through 10-fold cross-validation. There was a tendency to achieve better cropland mapping performance in the dry season and peak growth cycle in all three sites. We showed that monthly and seasonal cropland mapping accuracy was only slightly (~2%) increased using training data in corresponded month. Therefore, we recommended transfer seasonal/yearly training data to fine temporal resolution (i.e., monthly and seasonal) cropland mapping for time and labor efficiency.