时间序列案例推理检测土地利用短期快速变化

在实际的地理环境中,不同的土地利用/土地覆盖类型往往具有特定的时/空变化特征,适合于采用案例匹配方法进行求解。但目前常用的案例推理多以静态推理为主,缺乏对动态变化过程的推理。论文利用案例的空间可重用性,提出了基于时间序列的案例推理方法,通过分层随机采样挑选案例点,建立了试验区同步观测案例库,并对2005～2006年试验区15景Radarsat图像进行了时间序列影像分割,利用案例点的时间序列特征值对分割后的图斑进行时间序列相似性测度,生成时间序列案例匹配矩阵,并进一步得到变化图斑的土地利用转化类型和变化时间。最后,结合野外实测资料,对案例匹配结果进行了误差评价,在所有地类变化检测结果中,菜地由于种植类型多样,物候差异大,变化检测误差最大,其次是平整土地;鱼塘(11.3%～13.2%)和果园(10.7%～16.9%)变化检测误差最小。和基于规则的变化检测方法相比,案例推理方法精度略高。

Land-use Short-term Change Detection Based on Time-series-case-based-reasoning(TSCBR)

Remote sensing is one of the most efficient ways of land-use/land-cover monitoring.So far most land-use data could only be updated every one or two years by satellite.Recently,short-term land-use/land-cover change monitoring is gaining increasingly concern in urban planning, land management and ecological science.While comparing to long-term change detection,short-term land-use/land-cover change detection is relatively more difficult.In short-term change detection,both the land-surface reversible seasonal change and the in-reversible land-use change,such as man-made changes are remarkable in difference scenes of images.Hence any feasible short-term change detection method should be able to tell the difference between seasonal land-cover change and land-use change.In this paper,a time-series-case-base-reasoning (TSCBR) method was proposed to trace the dynamic variation of local land-use/land-cover change,which can extend the CBR from temporal reuse to spatial reuse.And 15 scenes of Radarsat-1 fine images were taken for trail in the Pearl River Delta,south of China.Firstly,stratified sampling was carried out and case parcels of different classes were selected.Secondly,time series image segmentation was performed on the 15 scenes of time-series images and backscatters and texture characteristic were extracted basing on the segmented objects.Then objects-based similarity comparison was performed between cases parcels and other parcels in the study area.k-NN algorithm was adopted and the most of the matched case types were assigned to the parcel compared.For each parcel a time series match matrix was generated and the parcel whose matched-case type varied from one type to another was defined as changed parcel.Finally,field collected data was used for validation.The classification error of TSCBR was about 20.5% and change detection error was about 11.8%.In all the classes,vegetable plot gains the maximum classification error (29.8%) and change detection error(15.2%). The next is development area(classification error 24.9%,change detection error 14.9%) and orchard(classification error 27.5%,change detection error 14.8%).Big error in vegetable plot owes mainly to its various plant pattern and fractal distribution in the study area.Besides,river gains the minimum classification error(7.9%) and change detection error(8.9%).In general,TSCBR proposed an applicable short-term changes detection method with only a small quantity of known cases. Comparing to traditional static case-base reasoning,TSCBR could trace the dynamic variation of land parcels in backscattering and image texture.This enable the real-time observation of short-term land-use change in fast developed area.