Local spectrum-trend similarity approach for detecting land-cover change by using SPOT-5 satellite images

Spectra-based change detection (CD) methods, such as image difference method and change vector analysis, have been widely used for land-cover CD using remote sensing data. However, the spectra-based approach suffers from a strict requirement of radiometric consistency in the multitemporal images. This letter proposes a new image feature named spectrum trend, which is explored from the spectral values of the image in a local geographic area (e.g., a 3 × 3 sliding window) through raster encoding and curve fitting techniques. The piecewise similarity between the paired local areas in the multitemporal images is calculated by using a sliding window centered at the pixel to generate the change magnitude image. Finally, CD is achieved by a threshold decision or a classified method. This proposed approach, called "local spectrum-trend similarity," is applied and validated by a case study of land-cover CD in Wuqin District, Tianjin City, China, by using SPOT-5 satellite images. Accuracies of "change" versus "no-change" detection are assessed. Experimental results confirm the feasibility and adaptability of the proposed approach in land-cover CD.