基于改进多元纹理信息模型和GF-1影像的县域冬小麦面积提取

作物面积监测具有较强的时效性,应用不断发展的遥感技术能够及时可靠地监测主要作物的种植面积。该文围绕国产高分一号卫星(GF-1)农业应用关键技术,研究县域尺度农作物种植面积快速准确提取的方法。在考虑多光谱遥感影像空间相关性的基础上,利用矩阵分解和距离空间转换等数学工具设计一种改进多元纹理信息(modified multivariate texture,MMT)提取模型,实现基于GF-1遥感影像的改进多元纹理信息提取、纹理与光谱信息融合以及基于融合影像分类的县域尺度冬小麦识别和面积提取。选用冬小麦出苗和越冬2期GF-1宽视场影像,结合地面实测数据和最佳识别时相遥感参考数据,对比分析基于光谱信息、单变量纹理与光谱融合信息、改进多元纹理与光谱融合信息的3种冬小麦识别和面积量算结果,实现对改进多元纹理信息效果以及小区域和较大区域上冬小麦面积提取的精度验证。试验结果表明:1)与基于其他2种传统分类特征信息的结果相比,改进多元纹理信息的加入能够显著提高冬小麦识别精度(出苗期提高4.12%和2.36%,越冬期提高2.59%和0.94%);2)在不考虑影像质量、生育期和地面样方测量精度等客观因素的影响下,基于该文方法的小区域内冬小麦面积提取精度普遍优于90%;3)在冬小麦长势稳定的时相(越冬期),基于该文方法的较大区域冬小麦面积量算结果能够达到接近最佳识别时相(孕穗期)的面积提取精度,二者的一致性程度超过97%。因而,利用GF-1宽视场影像能够有效提取县域尺度冬小麦种植面积,可为农作物监测业务运行提供遥感数据保障。

Area extraction of winter wheat at county scale based on modified multivariate texture and GF-1 satellite images

Winter wheat is grown in a wider area in China.Monitoring its planting area is therefore a key link for national food security.And area extraction is based on the availability of data source.As the first new satellite of GF series domestic satellites,GF1 satellite realizes the combination of high spatial resolution, multispectrum, and wide field of view (WFV) which has been applied in agricultural monitoring nearly three years.It is necessary to evaluated GF1 imagery for agricultural applications, especially for the planting area monitoring of food crops.In this paper, we studied the effectiveness of area extraction of winter wheat at a county scale using the WFV imagery from GF1 satellite.An approach using both spectral information and multivariate texture was proposed in order to make full use of spatial structure information in satellite imagery to further improve classification accuracy of the stable crop.Firstly, the multivariate texture was extracted through modeling based on the modification of multivariate variogram, and the model parameter measured spatial correlation with respect to all the bands of a multispectral image was designed as a distance metric computed by mapping the Mahalanobis distance between spatial point pair into the Euclidean space so that the Mahalanobis distance can be computed as Euclidean distance.We realized this transformation through the Cholesky decomposition of the covariance matrix item in the Mahalanobis distance expression.Then the derived multivariate texture image was combined with spectral data, and the fusion of spectral and texture information was input into the supervised classification technique of support vector machine to identify winter wheat.Two GF1 images acquired on November 2013 and January 2014 with four spectral bands (blue, green, red, and nearinfrared) and 16 m pixel size covering the large area of winter wheat in Suixi county in Anhui province were selected as remotely sensed data source.Classifications were generated based on the proposed method using the modified multivariate texture information and other two traditional methods using spectral data alone and plus traditional texture images at bitemporal, respectively.Accuracy assessments based on test samples were used to evaluate classification results.Compared with those from other two traditional methods, classification results from the proposal method had a significant improvement with the overall accuracy of 4.12% and 2.36% in seeding stage, and with the overall accuracy of 2.59% and 0.94% in overwintering stage.Ztest statistics were computed for paired comparisons among three classification results at bitemporal respectively, and the confidence level values were all less than 2.5%.Comparative analysis was carried out by computing the relative errors between the extracted area from classification results of the proposed method and the statistical area for winter wheat in some certain quadrats.With no consideration of the possible objective influence from imaging quality, phenology or measurement precision of quadrats, accuracies of the extracted area from classification results of the proposed method in quadrats were generally better than 90%.Then a Rapideye image covering the same area acquired on April 2014 with a higher resolution was used as a reference to evaluate the availability of the proposal method in the larger test regions with size of 5 km × 5 km.Consistency of the extracted area of winter wheat from classifications using GF1 images and the reference data from Rapideye image in test regions were able to achieve nearly 97% when winter wheat grown stably.From the results of experiments, it was found that fusion of multivariate texture and spectral information with GF1 satellite images could significantly improve the overall accuracy of winter wheat identification, and the proposed approach presented in this study also can be useful for different crops in other regions.