基于特征优选决策树模型的河套灌区土地利用分类

为了提高土地利用遥感识别精度，探索不同识别期及不同特征变量对土地利用类型遥感识别精度的影响。该研究采用Landsat时间序列影像数据，考虑不同月份和不同特征变量（波段、光谱指数及纹理特征）组合方式建立土地利用决策树分类模型，并利用河套灌区永济灌域实测数据和Google earth影像对不同组合方式的土地利用模型进行数量结构和空间布局的验证，筛选出最优的土地利用遥感模型并确定最佳识别期。结果表明：在不同月份Green（绿波段）和Ent（熵Entropy）分别与波段和纹理特征变量中的因子所含有的信息重复率最高，需剔除，归一化植被指数（Normalized Differential Vegetation Index, NDVI）和增强型植被指数（Enhanced Vegetation Index，EVI）在今后的研究中可选其一应用；与单一特征变量相比，不同特征变量组合后能提高模型精度，平均总体精度和Kappa系数分别提高了6.72个百分点和0.09。采用8月影像数据构建的遥感模型精度最高，最优遥感模型的特征变量组合方式为波段+光谱指数+纹理特征，总体精度、Kappa系数、制图精度和用户精度分别为80.23%、0.74、80.95%和86.26%，且减少了未利用地和居民工况用地空间布局的错分。通过综合比较，该研究区土地利用最佳识别期为8月，其次为9月。利用8月最优遥感模型（最佳识别期和最优组合）识别的耕地、林地、草地、未利用地、水域和居民工矿用地的制图精度分别为96.83%、73.33%、70.00%、65.52%、100.00%和80.00%，用户精度分别为76.62%、100.00%、82.35%、82.61%、100.00%和80.00%。因此可选用8月最优模型应用于长时间序列的土地利用类型识别。

Classification of land use in Hetao Irrigation District of Inner Mongolia using feature optimal decision trees

Up-to-date classification of land use types has become a critical component in current strategies to manage natural resources and the regional environment. Alternatively, remote sensing has also been widely used over the past 20 years as an effective tool for spatial data acquisition, particularly for the sustainable management of natural resources and economical perspective to the land use and land cover changes. However, the land use classification using remote sensing is subjected to the characteristics of dispersion and fragmentation in the Hetao irrigation district of northwest China in recent years. This study aims to quantify the effects of duration and characteristic variables on the recognition accuracy of remote sensing for land use types. A decision-tree model was also established to classify the land use types using the integrated band reflectance, spectral index, and texture feature of different periods based on Landsat time-series image data. The model was finally verified by the measured data and Google Earth images from the quantitative structure and spatial layout. The specific procedure was as follows. Firstly, the characteristic variables were extracted from the Landsat time-series images of different periods, including the features of band, spectra, and texture. Principal component analysis (PCA) was selected to extract the feature factors. Only a few independent variables were selected from multiple variables or factors, aiming to fully reflect the information of more original indexes. Secondly, seven schemes were constructed using the characteristic factors, including three single-category schemes (Scheme 1 to 3), and four combined-category schemes (Scheme 4 to 7). Finally, a classification model of land use was constructed and then verified in different periods via the decision tree. The results showed that: 1) The highest repetition rate was found in the Green and Ent (entropy) with other factors in different months. The correlation between NDVI and EVI was much higher to be selected in future research. 2) The combined feature variables greatly improved the accuracy of classification, where the average overall accuracy and Kappa coefficient increased by 6.72% and 0.09, respectively, compared with the single feature variable. 3) There were some effects of different recognition periods on the accuracy of the model. The accuracy of the classification model in the band, spectral index, and texture feature using remote sensing images in August was better than that of other periods, where the misclassification was reduced on the spatial layout of unused and residential land. Specifically, the overall accuracy, Kappa coefficient, producer accuracy, and user accuracy were 80.23%, 0.74%, 80.95%, and 86.26%, respectively. Correspondingly, the best identification period was August in the study area, followed by September. 4) The optimal remote-sensing model was utilized to identify the agricultural land, forest, grassland, wasteland, water bodies, and build-up land under the optimal recognition period and combination, where the high accuracies were achieved: 96.83%, 73.33%, 70.00%, 65.52%, 100%, and 80.00%, respectively. In addition, the user accuracies were 76.62%, 100%, 82.35%, 82.61%, 100%, and 80.00%, respectively. In a word, the feature optimal decision-tree model under the optimal identification period significantly reduced the amount of data and the difficulty of model application, particularly suitable for the long-time and spatial changes of land use types. The finding can provide promising technical support to effectively improve the accuracy of land use classification in modern resource management.