Temporally Consistent Forest Stand Segmentation Using Landsat Imagery

The object-based image segmentation techniques are widely utilized in environmental disciplines to partition remotely sensed imagery into objects representing distinct conditions, such as vegetation structure or landform. However, most approaches are applied to a single temporal snapshot, limiting their ability to update polygons over time. To address this, we proposed a temporally consistent segmentation algorithm based on a two-phase region growing approach designed to be applied to time series of annual Landsat surface reflectance composites. We developed and demonstrated this new approach over six fire-disturbed forested study areas in British Columbia, Canada, to dynamically delineate polygons over time as they underwent land cover change. Our approach maintained the existing boundaries for forest polygons with no land cover change while updating those subject to change as forest regenerated and followed successional processes. Rapidly recovering areas, such as Cariboo and Fraser-Fort George, showed increases in mean segment area from 12 to 21 and 14 to 25 ha, respectively, approaching or exceeding predisturbance values. Additionally, segment shape complexity increased over time, reflecting the structural development of recovering stands. This work demonstrated the potential of utilizing Landsat surface reflectance data to update forest polygons over time with reference to forest development and increasing maturity.