地块破碎度对软硬变化检测法识别冬小麦分布精度的影响

软硬变化检测作物识别(soft and hard change detection,SHCD)是一种新型的作物识别方法。该研究针对不同耕地地块破碎程度的农业景观地区进行SHCD冬小麦识别,分析地块破碎程度对SHCD冬小麦识别精度的影响。试验结果表明,在种植地块破碎试验区,SHCD的RMSE对分辨率不敏感,均小于0.15,bias也比较小,R2随着检测窗口的增加,相关性逐步升高,达到98%以上。在种植地块规整试验地区,亦能够得到相同试验结论。SHCD方法综合了硬变化(hard change detection,HCD)和软变化(soft change detection,SCD)各自的优势,能够达到稳定且较高的识别精度,不受影像分辨率的影响;有效地解决了SCD在硬变化区(纯净像元)受到光谱不稳定性和HCD在软变化区(混合像元)识别为"0-1"排他性结果的不足,保证了冬小麦的识别精度,为大范围进行冬小麦识别以及其他作物的变化检测识别提供前期的试验基础。

Impact of land fragmentation on identification of winter wheat distribution accuracy by soft and hard change detection method

Soft and hard change detection method(SHCD) is a newly proposed approach and previous studies have showed that this method is useful for accurately identifying crop. In this paper, SHCD was used for classifying winter wheat in both simple, homogeneous, low fragmented regions and complicated, heterogeneous, discontinuous regions, and the impact of agricultural landscape pattern and image resolution on the accuracy of winter wheat identification was quantified. Experimental process included simulation image creation, winter wheat mapping by SHCD and result analysis. Simulated images were obtained by the crop change detection model, the winter wheat phenology and the effects of parcel fragmentation. Winter wheat mapping was obtained by the processing of image differencing, sample selection, and ESVM division. Three statistical methods were used to estimate the precision of winter mapping for different window sizes. And we further analyzed the effects of image resolution, window size and spatial characteristics on the identification of winter wheat distribution accuracy. The results showed that: 1) The optimum resolution was 10 and 40 m for hard change detection(HCD) and soft change detection(SCD), respectively. Different from HCD and SCD methods, SHCD was not sensitive to pixel resolution and always yielded accurate classification results. 2)From the view of efficiency, the calculating process including hyperplane segmentation and labelling was same for SHCD, SCD and HCD methods. During the labeling stage, SCD directly assigned the membership probability to each test pixel, while HCD segmented the feature space using hyperplane and adopted thresholds to determine the label for test pixels. By combining the advantages of HCD and SCD, SHCD firstly segmented the feature space into wheat, mixed wheat and non-wheat area using marginal hyperplanes, and then assigned the membership probability for mixed wheat. Accuracy assessment results showed that: 1)In highly fragmented regions, the root mean square errors(RMSEs) of SHCD were lower than 0.15 and not sensitive to image resolution. The bias values were low; the R² values were higher than 98% and increased with window size increasing. 2)In lowly fragmented regions, SHCD was also not influenced by image resolution. Compared with SCD and HCD, wheat distribution derived from SHCD also had the lowest RMSE and bias and the highest R² value. Combining the advantages of HCD and SCD method, SHCD method effectively eliminated the classification errors caused by spectral variability in hard change areas by SCD method and exclusive result in soft change areas by HCD method. Therefore, SHCD method provided the highest accuracy of crop acreage. So this simulation experiment provides experimental basis and ideas for winter wheat identification in real situations. However, SHCD is limited to classification errors of ESVM method, which is difficult to identify crop from other land cover types between land parcels. Besides, the impact of the different land cover types with similar spectral characteristics is inevitable for this method. Finally, even though SHCD shows good results for soft change area, the errors of commission are still large, which needs to be solved in the future. At the same time, in order to reduce the impact of other factors, simulation experiment simplified real situations, which needs to be tested in real research areas furthermore.