Trend forecast based approach for cropland change detection using Lansat-derived time-series metrics

Accurate information on cropland changes is critical for understanding greenhouse gas emissions, biodiversity, food safety, and human welfare. Traditional bi-temporal change detection methods using remotely sensed imagery may generate pseudochanges due to phenological differences and interference factors. In this study, we develop the Trend Forecast-based change detection approach (TFCD) using Landsat-derived time-series metrics to eliminate pseudochanges caused by phenological differences. Assuming that time-series images could be modelled and analysed, the time-series model would have a high capacity for revealing trends and temporal patterns. The spectral variability of cropland has strong seasonal dynamics, which shows short-period regular changes and long-term dynamic trends. Therefore, multi-harmonic model is used to describe the trend and temporal patterns of cropland over time. Then, the differences between model predicted and observed trajectory are used to detect the change areas. Finally, the change types are determined using the model coefficients. The effectiveness of this method was verified using a stack of (25 images) Landsat Enhanced Thematic Mapper Plus and Operational Land Imager images from two years (2014 and 2015). The results indicated that TFCD correctly detected true changes, with 95.79% overall accuracy and a Kappa coefficient of 0.751, and that the method was superior to the traditional methods.