A cultivated-land non-agriculturalization remote-sensing monitoring method based on multi-scale data spatiotemporal assimilation and time–space–spectral feature fusion

Abstract

The present invention relates to a cultivated-land non-agriculturalization remote-sensing monitoring method based on multi-scale data spatiotemporal assimilation and time–space–spectrum feature fusion, and belongs to the field of multi-scale remote-sensing monitoring of cultivated-land use-type change. Aiming at the difficulty of acquiring change-label samples, the method first performs accuracy optimization on existing land-cover classification products, and automatically acquires large-scale cultivated-land change-detection training samples using an evolved training-sample generation method. Landsat and Sentinel-2 remote-sensing imagery are jointly used to reconstruct high–spatiotemporal-resolution time-series data, and a time–space–spectrum multidimensional continuous-change detection technique is adopted to accurately extract change breakpoints and obtain high-accuracy attributes of change time, location, and extent. A land-cover classification model based on ensemble learning is constructed, trained with the generated large-scale training samples, and the trained model is then used to detect remote-sensing imagery before and after the breakpoint to identify change type and change direction.

Images