JIA-2019-11

2636 ZHANG Xi-wang et al. Journal of Integrative Agriculture 2019, 18(11): 2628–2643 In another validation method, these interpreted validation points are considered to be the true values. As is defined as the percentage of correctly identified validation points: As (%)= ×100 ∑ ∑ A R Sample Sample (11) where ∑Sample R is the number of correctly identified validation points; and ∑Sample A is the total number of validation points. A confusion matrix is applied in the identification result assessment. Overall accuracy (OA), producer’s accuracy (PA), user’s accuracy (UA) and the kappa coefficient (Kappa) are calculated to analyze the winter wheat identification effect. The root mean square error (RMSE) is also used to evaluate the accuracy of the identified area at the scale of each county. RMSE is defined as: ( ) ∑ − n i =1 i 0 i / n AA 2 RMSE= (12) where n is the number of sub-regions (counties), A i is the identified acreage of winter wheat in the i th sub-region, and A i0 is the reference acreage. 4. Results 4.1. Selection of temporal change features The NDVI values at pointsA, B, C, D, E, and F are calculated based on eq. (3) from the MODIS NDVI time series data. A correlation analysis is conducted between the slopes and abundance samples, and the results are shown in Table 1. As shown in Table 1, the maximum correlation coefficient 0.8999 is between |Slope CD | and abundance samples, proving that |Slope CD | is an important indicator for abundance assessment. |Slope EF | and |Slope AB |, with the correlation coefficients 0.8156 and 0.7391, respectively, are also good indicators. However, |Slope BC | and |Slope DE | are almost irrelevant to abundance, with the correlation coefficients –0.0077 and –0.2463, respectively. Based on these results, we selected |Slope CD |, |Slope EF | and |Slope AB | to construct models. 4.2. Abundance assessment models To assess winter wheat abundance, multiple regression analyses were performed between NDVI time series curve slopes and the acreage fractions of 56 winter wheat abundance samples according to the spatial locations of data plots. Regression models were established as shown in Table 2. Clearly R 2 reaches or exceeds 0.8098 only when |Slope CD | is used for the regression analysis. This result further demonstrates the importance of CD for abundance assessment. In the multiple regression model using the combination of |Slope CD | and |Slope AB |, R 2 reaches 0.8146. R 2 reaches 0.8158 using the combination of |Slope CD | and |Slope EF |, illustrating that goodness of fit can be improved when either |Slope AB | or |Slope EF | is added. Using the combination of |Slope CD |, |Slope AB | and |Slope EF |, R 2 is not significantly improved compared to the previous models, since |Slope AB | and |Slope EF | have a very high correlation coefficient, 0.9264. However, the R 2 of this three-component multiple regression model is the largest at 0.8159. 4.3. Abundance of winter wheat According to the analysis in Table 2, the three-component Table 1 Correlation matrix between abundance samples and each slope Variable Abundance |Slope AB | |Slope BC | |Slope CD | |Slope DE | |Slope EF | Abundance 1.0000 |Slope AB | 0.7391 1.0000 |Slope BC | –0.0077 –0.2257 1.0000 |Slope CD | 0.8999 0.7722 –0.0967 1.0000 |Slope DE | –0.2463 –0.4460 0.3623 –0.2332 1.0000 |Slope EF | 0.8156 0.9264 –0.3300 0.8627 –0.3401 1.0000 Table 2 Regression models between abundance samples and each slope 1) Dependent variable Argument Regression models R 2 F -value P -value y x 1 y= 2.6116 x 1 +0.7076 0.5436 65.02 7.79E-11 y x 2 y= 4.4051 x 2 –0.4656 0.8098 229.89 4.14E-21 y x 3 y= 1.6125 x 3 +0.2921 0.6651 107.27 1.95E-14 y x 1 , x 2 y= 0.3869 x 1 +3.9912 x 2 –0.3460 0.8146 116.45 4.01E-20 y x 1 , x 3 y= –0.4082 x 1 +1.8241 x 3 +0.2332 0.6670 53.09 2.20E-13 y x 2 , x 3 y= 3.7574 x 2 +0.3032 x 3 –0.3641 0.8158 117.37 3.39E-20 y x 1 , x 2 , x 3 y= 0.1093 x 1 +3.7734 x 2 +0.2410 x 3 –0.3511 0.8159 76.84 4.10E-19 1) x 1 is |Slope AB |, x 2 is |Slope CD |, x 3 is |Slope EF |.