基于物候特征的盐渍化信息数据挖掘研究

盐渍化是影响植被和作物长势的重要因素,精确反演盐渍化的时空分布信息至关重要。基于MOD13A1-NDVI数据反演生长季开始日期(SOS)、生长季结束日期(EOS)、生长季长度(LEN)等物候参数和计算出能高精度反演盐渍化空间分布的多种植被指数、盐分指数、地形指数、干旱指数等参数后作为BP-ANN人工神经网络的输入因子来反演盐渍化信息,同时按照植被类型和地貌类型进行分区来反演盐渍化信息,以探讨盐渍化受植被和地貌类型的影响。主要结论如下:(1)盐渍化的形成受多种因素的影响,与物候参数大多呈非线性关系,不能单纯的以某拟合公式来进行表达,需要借助人工神经网络超强的非线性拟合能力来反演盐渍化信息。(2)通过深入挖掘植被物候信息,在融入物候参数后的反演精度显著提高。可决系数R2从0.68(非物候参数)增加到0.79(包括物候参数),但是需要加入地形、影像数据和土壤水分等方面的信息来更加精确的反演盐渍化信息。生物累积量指标LSI(Large seasonal integral)和SSI(Small seasonal integral)能够很好的表征盐渍化的信息。(3)划分植被类型后的盐渍化提取精度进一步提高,可决系数R~2达到了0.88。(4)以地貌特征作为类型分区后,反演结果的R~2达到了0.85,精度较高,比以植被类型作为分区的精度略小。高程较低区域的盐渍化现象普遍较重,盐渍化程度受到地形和地貌因素的影响显著。(5)农用地区域多为非盐渍化和轻度盐渍化地,稀疏植被区多为重盐渍化地。研究区的非盐渍化和轻盐渍化地、中盐渍化地和重度盐渍化地比例分别为53.42%,13.71%,32.87%。以上的研究结果提出了一种融合物候信息和非物候参数来反演盐渍化信息的方法,进行深入的协同植被物候监测盐渍化信息方面的数据挖掘,在融入了物候参数后,盐渍化的预测精度显著提高。

Research on data mining of salinization information based on phenological characters

Soil salinization is an important factor that affects crop and vegetation growth condition and can result in environmental impacts with considerable economic consequences. Therefore, it is necessary to determine an effective method to monitor spatiotemporal salinity distribution. We used MOD13A1 time-series NDVI data to determine the vegetation phenology, including start of season (SOS), end of season (EOS), length of season (LEN), etc., and calculated several vegetation, salinity, terrain, and drought indexes, and spatial models. These were used as input parameters for the BP-ANN model. Meanwhile, we predict the soil salinity through vegetation and geomorphological partitioning, which described the correlations between vegetation or geomorphic type and salinization. The main conclusions are as follows: salinity is influenced by many factors, and many of them show non-linear relationships between phenological indicators and salinization, so we utilized artificial neural networks to predict soil salinity than mathematical equations; through a combination of phenology parameters, the precision of inversion salinity R² improved from 0.68 (no phenologcial indicators were included) to 0.79 (phenological indicators were included). However, additional auxiliary data to predict soil salinity, such as terrain, image, and soil moisture parameters should also be included. After the classification of the vegetation, the inversion precision improved obviously, where R² increased to 0.88. Phenological characters, such as large seasonal integrals (LSIs) and small seasonal integrals (SSIs) are good indicators to represent soil salinity. After geomorphological partitioning, R² increased to 0.85, indicating that it could be a good salinity predictor, but the ability of comprehensive inversion was lower than vegetation type partitioning. In farmland, the salinity level was low. The low, intermediate, and high salinization was 53.42, 13.71, and 32.87% respectively. Generally, salinization was higher at lower altitudes, and the salinity level was affected by terrain and geomorphological factors. The above conclusions indicate an effective method for the inversion of salinization levels that combines phenology and other parameters for comprehensively determining the effect of phenological information on salinity monitoring ability in data mining. The inversion of soil salinity is enhanced by the inclusion of phenological parameters.