Study On Remote Sensing-based Drought Monitoring Method Derived From Multi-Source Data

Drought is a complex natural hazard, and it is affected by many factors. Climate change brought about by global warming causes the intensification and increased frequency of droughts, which result in considerable impacts and economic losses. A comprehensive study on drought monitoring and prediction is important to provide policymakers with accurate and timely information needed for early warning and mitigation. Based on the summarization of previous research work on remote sensing drought monitoring, this paper mainly focus on the remote sensing-based drought monitoring method derived from multi-source data. Evolution of drought and wetness has been analyzed for the period of 1961-2012 using in situ reference data. To explore the optimal drought method, this paper attempt to study drought based on the analysis of shortwave infrared spectral feature space and drought-related information mining and in situ observation data are used to evaluate the performance of the proposed model. The main conclusions are as follows:(1) The Standardized Precipitation Index (SPI) and the Standardized Precipitation-Evapotranspiration Index (SPEI) on various time scales were used to study drought evolution in China from 1961 to 2012. Results showed that the spatial patterns of drought severity change implicated by SPI and SPEI were basically alike and these two indices could characterize regional drought conditions effectively. Additionally, the SPEI identified a greater severity of drought than that identified by the SPI, largely because the SPI values are based only on precipitation data while the SPEI is based on both precipitation and temperature data and the very warm temperatures during those decades may have an influence on the SPEI values. The non-parametric Mann-Kendall trend test was used to detect the trend of in situ reference data. Results showed that 1) a downward trend was detected at 38 stations at high confidence levels and a nonsignificant downward trend was detected at 277 stations for average precipitation; 2) most of the stations showed an increasing trend in mean temperature at high confidence levels; 3) most of the stations exhibited significant drying trend (in terms of the SPEIs) and these stations were mainly located in Northern China (Inner Mongolia, Hebei and Shanxi), Northeast China (Liaoning, Jilin and Heilongjiang), Northwest China (Shaanxi, Gansu, Ningxia, Northern of Qinghai and Southeast of Xinjiang) and Southwest China (Southeast of Sichuan, Yunnan, Chongqing and Guizhou).(2) Changes in water content in plant tissues or soil have a large effect on the reflectance in several regions of the shortwave infrared spectrum. A new remote sensing-based drought index, the MODIS shortwave infrared water stress index (MSIWSI) was proposed for monitoring agricultural drought, by integrating the information from the shortwave infrared channel 6 and 7 from the MODIS satellite data based on the distribution of different land cover classes in MODIS shortwave infrared spectral feature space. Correlation and regression analyses were performed between MSIWSI along with two other well-known drought indexes, namely, the enhanced vegetation index (EVI) and the Modified Perpendicular Drought Index (MPDI) and ground-measured soil moisture data to assess the capability of MSIWSI over arid regions in northwest China. Results showed that MSIWSI presented a better performance than EVI and MPDI for soil moisture retrieval. Additionally, the spatial drought conditions of the MSIWSI maps were compared with the actual drought intensity using the statistical data of agricultural meteorological disaster and the temporal distribution of the precipitation and mean temperature data at the agro-meteorological sites. The practical application of MSIWSI in northwest China also demonstrated that it could provide accurate and detailed drought condition.(3) The model input data was selected based on correlation analyses between remote sensing drought index and in situ variables. The Synthetic Minority Over-sampling Technique (SMOTE) method was used to balance imbalanced training datasets. Three typical classifiers:Back-Propagation Neural Network (BPNN), Support Vector Machines (SVM) and Classification and Regression Trees (CART) were applied for classification of regional drought grade. The results showed that 1) BPNN achieved more accurate classification results than SVM and CART in training dataset and testing dataset; 2) the SVM classifier was more efficient than the CART classifier in training dataset, while the CART classifier produced a more accurate classification than the SVM classifier in testing dataset; 3) the capability of each classifier in drought grade classification varies along seasonal time. Considering the limitation of single classifier, three classifier ensemble methods:Majority Voting (MV), Fuzzy Majority Voting (FMV) and Fuzzy Majority Voting with a Threshold (FMVT) were introduced to fuse the three single classification results. Experimental results clearly demonstrated that 1) all three ensemble methods could improve overall classification accuracy in both training dataset and testing dataset; 2) FMVT ensemble method performed the highest overall accuracy in testing dataset, which was respectively 3.4%,5.6% and 4% higher than BPNN, SVM and CART classification. Additionally, compared with three single classifiers, FMVT ensemble method achieved more accurate classification results at all different time periods, indicating that FMVT ensemble method can be effectively used for classification of regional drought grade.(4) The optimal drought-related factors over different time periods were selected through correlation analyses between remote sensing drought index and in situ variables. Based on the Random Forest (RF) technique, we established a new remote sensing index named Integrated Drought Condition Index (IDCI), which considers land cover data, climate classification information, digital elevation data, and multisource drought-related factors comprehensively. The correlation coefficients and RMSE values were calculated between the remotely sensed drought indices (VCI, TCI, PCI, SMCI, VHI and IDCI) and the 3-month SPEI. Results showed that IDCI produced the highest correlation coefficient values with in situ variables and all the correlations passed the p value<0.05 significance test. Additionally, regression analyses were performed between the IDCI and the in situ reference data to further evaluate the capability of regional drought condition monitoring. Analyses were performed in nine main provinces of the study area. Results showed that the IDCI agreed well with the SPEI-3 in different provinces, and all the correlations were statistically significant (p<0.005). The yearly IDCI variations in 21 representative meteorological sites were compared with that of the in situ drought indices to evaluate the temporal drought monitoring capability of this index. Results showed that the IDCI exhibited consistent variations with the in situ reference data at the regional scales in most cases. The spatial changes in the IDCI maps were also compared with the changes in the in situ reference data to assess the IDCI performance in monitoring short-term drought conditions. Results showed that the two variables basically showed a similar spatial pattern. We conclude that IDCI can be used to characterize drought conditions and patterns effectively.