Object-based Image Analysis With Machine Learning Algorithms For Cropping Pattern Mapping Using GF-1/WFV Imagery

Cropping pattern is a fundamental description of land cover/use and is related to a number of ecological factors and the regulation of agricultural management. Satellite-based earth observation provides an ideal basis for the area-wide and spatially detail in crop distribution. Remote sensing potential for monitoring crops using high-resolution images is more recently discussed. However, there are some limitations in currently data and method: 1) The potential to characterize the spatial pattern of main crop types in a spatially detailed and comprehensive manner at local and regional scales is limited by popular satellite data, and the new remote sensing imageries are paid insufficient attention to. 2) A trade-off relationship of spatial, spectral and temporal resolution limits in data mining and deep learning. 3) Classification strategy is difficult to provide timely and accurate information for warranting a sustainable development or implementing effective adaptation and mitigation strategies for policy making. To solve these problems, the study mainly used GF-1/WFV imageries combining object-based imagery analysis with machine learning algorithm to explore the potential and valuable application of spectral-textural features and multi-season classification strategy. The conclusions of this study are summarized as follows:(1)A literature review suggests that the medium and low resolution satellite imageries are widely used in crop pattern mapping, especially MODIS and Landsat TM/ETM+. However, the potential application of high resolution satellite data for crop mapping has been the important research topics. A good quality/cost relationship and a proper combination of spatial, spectral and temporal resolution is a prerequisite for agricultural land use information.(2)Object-based crop pattern mapping can deeply mine temporal and spatial characteristic in high resolution satellite imageries. After testing different parameter values and qualitatively analyzing the multi-resolution segmentation results based on GF-1 WFV sensor at a 16-m resolution in Beian which is the key monitoring area of expanding corn and soybean rotation, and named as “Sickle bay”, it was decided to adopt the following segmentation parameters: scale parameter 30, shape 0.1 and compactness 0.5. The basic idea of object-based image analysis is the segmentation of the image and the construction of a hierarchical network of homogenous objects that match to plot boundaries to overcome the problems due to pixel heterogeneity and crop variability within the field.(3)The mono-temporal GF-1 WFV imagery was used to investigate variable importance and mapping accuracy’s change over the growing season. To assess the effect of the mtry and ntree parameters on classification accuracy, numerous trees were created for various mtry and ntree values. Considering the computation time of the RF, the ntree and mtry parameter was set to 1000 and 5. When selecting the optimal temporal window, we first selected the single time period with the lowest OOB error. MDG is used to assess the variable importance in single time period. The spectral properties of the crops and the plant canopy and textural properties of the crop structure can character temporal and spatial variations of crop. In addition, variable importance derived from near-infrared was greater than that of other variables in the single time period. The overall accuracy of crop pattern mapping can be improved by 2.84% by adding the textural properties.(4)Classification strategy is optimized in way of importing multi-temporal spectral and textural properties in the form of month by month. MDA is used to assess the variable importance in multi-temporal time series. Variable importance derived from near-infrared was greater than that of other variables in time series. The contribution of textural properties in the feature space which is lack of spectral properties is higher than that in the feature space which is abundant spectral properties. The exhaustive method is adopted to verify the feasibility of classification strategy due to low misclassification error. The number of crop types is gradually increase and the overall accuracy is 91.33% in four temporal combination.(5)An accuracy assessment was performed for SVM classification and Landsat produced in this study to evaluate how well predictions based on different classifier and data source. It is found that both SVM and RF algorithms produced similar classification accuracies. SVM based model achieved higher classification accuracies than those produced by RF when the volume of feature space is small. But RF performed better than classifications produced using SVM in high dimensional feature space. The generalization ability of bivariate classification types derived from SVM was stronger than that of RF. It is opposite in many classification types. In addition, the crop pattern mapping results based on spectral and textural features from GF- 1 satellite data are better than Landsat.