Optimizing the feature variables for irrigated farmland mapping

Irrigation has been one of the most important land management in modern agriculture.Accurate mapping of irrigated arable land can provide more available data for study on food security, water resources, and climate change. Among them, the selection of feature variables is one of the most important steps in the process of irrigated farmland mapping. Therefore, this study aims to optimize the feature variables for mapping an irrigated farmland using the spatial distribution map and irrigation information data. Nebraska State in America with an excellent irrigation database was taken as the research area. The samples were first extracted from the irrigated and rain-fed farmlands in the database. Four types of 82 feature variables in the samples were calculated, including the monthly mean of precipitation, Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Greenness Index (GI), Normalized Difference Water Index (NDWI), daily Land Surface Temperature (LSTday), night Land Surface Temperature (LST night), the Land Surface Temperature difference between day and night (LSTdifference), Crop Water Deficit Index (CWDI), and Crop Water Stress Index (CWSI), while, the total precipitation, the mean NDVI, NDWI, LSTday, LSTnight, LSTdifference, CWDI, CWSI, as well as the Irrigation Probability Index (IPI), and Water-adjusted Green Index (WGI) in the growing season. Random forest was utilized to determine the importance of 82 feature variables to the identification of irrigated farmland. The results showed that the contribution to the identification of irrigated farmland was ranked in the order of the comprehensive > vegetation > soil > meteorological feature variables. As such, the 16 best feature variables were selected, including eight comprehensive, seven vegetations and one soil feature variable, but there was no meteorological feature variable. The CWSI, IPI, vegetation index, and LST difference were the sensitive characteristic variables to distinguish the irrigated farmland from the rain-fed farmland. There were also some differences in the best phase to identify the irrigated farmland with different feature variables. There was high sensitivity to irrigation for the CWSI in almost every month and the whole growing season. In the vegetation index, the more sensitive phase to distinguish the irrigated farmland from rain-fed farmland was concentrated in the later stage of the growing season. In the LSTdifference, September was the most sensitive month to distinguish the irrigated farmland from rain-fed farmland. The top four feature variables of importance ranking included the CWSI in April and May, the EVI in July, and IPI in the growing season. There was the highest overall classification accuracy (88.44%) for the first 16 important feature variables. Consequently, it infers that the remote sensing classification features have a great impact on the recognition accuracy of targets to be classified. The finding can also provide a strong reference for the selection of feature variables in the follow-up research on irrigation farmland mapping.