Remote sensing evaluation of drought degree based on threshold-optimized fuzzy majority voting model

Abstract: Drought affects not only agriculture, but also triggers negative economic, social, and environmental impacts. This study proposes a multiple classifier fusion method, threshold-optimized fuzzy majority voting (TFMV), for agricultural drought category evaluation. The standardized precipitation index (SPI) can be flexibly designed to measure drought severity for a specific time period. The 3-month SPI (SPI-3) was computed based on the long-term monthly precipitation record. Considering that the relationship between different remote sensing drought indices and SPI varies over time, the correlation coefficients were calculated between the remote sensing drought indices of each month from April to October, and SPI-3 from 2003 to 2012 to selected the input data of model. The results showed that the correlation coefficient values between the vegetation-related indices and SPI-3 varied over the different time periods. The VCI(vegetation condition index) showed the highest correlation with the SPI-3 in August because of the vegetation phenological phase. The correlation coefficient between the TCI(temperature condition index) and SPI-3 were statistically significant (P<0.01) except in May. All the correlation coefficient between soil moisture-related drought indices and SPI-3 were statistically significant (P<0.01) and all the correlation coefficients were above 0.45. The precipitation-related indices with various timescales showed high correlation coefficient with the in situ drought index, and these indices mostly have the highest correlation coefficient values with in situ drought index in the same timescales as that of the precipitation-related indices. Based on correlation analysis between remote sensing drought index and in situ drought index over the different time periods, VCI, TCI(temperature condition index), SMCI(soil moisture condition index)) and PCI-3 were selected as the input data of model. Since the distribution of the training data among draught classes is uneven, 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 assessment of regional drought category. The results showed that the capability of each single classifier in drought grade classification varies along seasonal time and the overall precision of these three classifiers for all samples from April to October were 69% (BPNN), 67.49% (SVM) and 69% (CART), respectively. Considering the limitation of single classifier, two classifier ensemble methods, majority voting (MV) and threshold-optimized fuzzy majority voting (TFMV) were introduced to fuse the three single drought category results. Experimental results clearly demonstrated that: 1) Ensemble method could improve overall classification accuracy; 2) TFMV ensemble method performed the highest overall accuracy in validation dataset, which was respectively 3.6, 5.1 and 3.6 percent point higher than that of BPNN, SVM and CART classification. Additionally, compared with majority voting method, TFMV achieved more accurate classification results in all different time periods. Additionally, the spatial drought conditions of the TFMV maps were compared with the actual drought intensity using the agro-meteorological disaster data recorded and the temporal distribution of the precipitation and mean temperature data at the agro-meteorological sites. Results showed that the TFMV maps exhibited consistent variations with the in situ reference data. The practical application of TFMV demonstrated that it can provide accurate and detailed drought condition and TFMV method can be effectively used for regional agricultural drought category evaluation.