Detecting Frequent Harvest of Alfalfa with Spatio-temporal Fused Data

With its high quality for the livestock industry, alfalfa is a valuable field crop planted globally. Since alfalfa can be harvested/cut multiple times in the growing season, the timing and frequency of alfalfa mowing events are crucial in forage management. Previous research suggested that remote sensing observations have considerable potential for monitoring forage mowing timing and frequency. However, single sensor data is often limited in detecting the frequent cutting of alfalfa due to its limited temporal resolution and the inevitable contamination of atmospheric conditions. This study aims to detect the mowing timing and frequency of alfalfa in Oklahoma with fused high spatial and temporal resolution data, which is derived by combining all available Landsat 8 OLI, Landsat 7 ETM+, and Sentinel-2 MSI optical imageries. We calculated the vegetation indexes (EVI and NDVI) and water indexes (NDMI and NDWI) for four typical alfalfa fields in Oklahoma, and combined the indices data from different sensors utilizing the virtual image pair-based spatio-temporal fusion (VIPSTF) method. The generated datasets had a 10-m spatial resolution and an average 6-day time interval of observation. A random forest (RF) model was subsequently trained and simulated the mowing events based on the datasets above. Results showed the model performed well with an accuracy of 0.94. The model was also applied with the Harmonized Landsat Sentinel-2 (HLS) data, and the comparison between the two datasets showed the VIPSTF fusion data with higher spatial and temporal resolution captured more complete harvesting information in the growing season (from April to October). This study confirmed the operability of machine learning models based on fused remote sensing data with a high spatial and temporal resolution for small-scale forage harvest detection, provided an effective remote sensing tool for precision forage management.