0 引 言
1 研究区域与数据集
1.1 研究区域
表1 农业干旱预测研究采样气象站点基本信息Table 1 Basic information of meteorological stations sampled for agricultural drought prediction research |
地点 | 经度/(°) | 纬度/(°) | 海拔/m |
---|---|---|---|
信阳 | 114.03 | 32.08 | 115.4 |
固始 | 115.37 | 32.10 | 43.9 |
阜阳 | 115.44 | 32.52 | 33.9 |
霍邱 | 116.47 | 32.33 | 64.4 |
定远 | 117.40 | 32.32 | 70.6 |
1.2 数据集构建
表2 农业干旱预测研究数据集划分Table 2 Dataset division for agricultural drought prediction research |
数据集划分 | 起始时间 | 结束时间 | 总数/条 |
---|---|---|---|
训练集 | 2000/1/1 | 2016/4/15 | 5 950 |
验证集 | 2016/4/16 | 2018/8/13 | 850 |
测试集 | 2018/8/14 | 2023/4/9 | 1 700 |
2 GCN-BiGRU-STMHSA模型
2.1 空间特征提取模块GCN-SMHSA
2.2 时间特征提取模块BiGRU-TMHSA
3 评价指标
3.1 农业干旱指标
表3 农业干旱预测研究SSMI指标分级表Table 3 SSMI index classification table for agricultural drought prediction research |
指标范围 | 分类 | 等级 |
---|---|---|
0<SSMI | 正常 | 1 |
-1<SSMI≤0 | 轻度干旱 | 2 |
-1.5<SSMI≤-1 | 中度干旱 | 3 |
-2.0<SSMI≤-1.5 | 重度干旱 | 4 |
SSMI≤-2.0 | 特旱 | 5 |
3.2 模型评价指标
4 结果与分析
4.1 实验结果分析
表4 在不同地区的SSMI预测与分类对比试验结果Table 4 Comparative trials of SSMI prediction and classification results in different regions |
模型 | GRU | LSTM | BiGRU | Transformer | GCN-BiGRU | GCN-BiGRU-STMHSA | |
---|---|---|---|---|---|---|---|
信阳 | MAE | 0.116 | 0.124 | 0.097 | 0.084 | 0.078 | 0.066 |
RMSE | 0.152 | 0.163 | 0.128 | 0.118 | 0.119 | 0.087 | |
R 2 | 0.724 | 0.683 | 0.804 | 0.834 | 0.737 | 0.857 | |
ACC | 0.874 | 0.863 | 0.894 | 0.918 | 0.861 | 0.925 | |
F 1 | 0.874 | 0.863 | 0.894 | 0.918 | 0.860 | 0.924 | |
阜阳 | MAE | 0.080 | 0.092 | 0.061 | 0.050 | 0.047 | 0.044 |
RMSE | 0.102 | 0.117 | 0.081 | 0.071 | 0.667 | 0.061 | |
R 2 | 0.778 | 0.711 | 0.859 | 0.893 | 0.877 | 0.882 | |
ACC | 0.886 | 0.871 | 0.912 | 0.925 | 0.924 | 0.900 | |
F 1 | 0.886 | 0.871 | 0.912 | 0.925 | 0.924 | 0.898 | |
霍邱 | MAE | 0.106 | 0.112 | 0.089 | 0.081 | 0.061 | 0.057 |
RMSE | 0.130 | 0.143 | 0.122 | 0.116 | 0.092 | 0.079 | |
R 2 | 0.730 | 0.714 | 0.792 | 0.813 | 0.839 | 0.860 | |
ACC | 0.845 | 0.833 | 0.884 | 0.887 | 0.924 | 0.888 | |
F 1 | 0.845 | 0.833 | 0.884 | 0.887 | 0.924 | 0.887 | |
固始 | MAE | 0.141 | 0.148 | 0.130 | 0.155 | 0.067 | 0.053 |
RMSE | 0.183 | 0.191 | 0.172 | 0.201 | 0.096 | 0.071 | |
R 2 | 0.592 | 0.555 | 0.641 | 0.506 | 0.827 | 0.880 | |
ACC | 0.793 | 0.776 | 0.809 | 0.770 | 0.911 | 0.925 | |
F 1 | 0.791 | 0.772 | 0.808 | 0.768 | 0.911 | 0.924 | |
定远 | MAE | 0.108 | 0.128 | 0.091 | 0.083 | 0.074 | 0.059 |
RMSE | 0.139 | 0.164 | 0.122 | 0.116 | 0.108 | 0.079 | |
R 2 | 0.756 | 0.661 | 0.813 | 0.831 | 0.802 | 0.884 | |
ACC | 0.834 | 0.794 | 0.861 | 0.877 | 0.899 | 0.913 | |
F 1 | 0.834 | 0.794 | 0.861 | 0.877 | 0.898 | 0.910 |
4.2 不同时间尺度实验结果分析
表5 GCN-BiGRU-STMHSA模型在不同时间尺度的SSMI预测及分类结果Table 5 Prediction and classification results of SSMI at different time scales using the GCN-BiGRU-STMHSA model |
预测步长/天 | 7 | 14 | 21 | 28 | |
---|---|---|---|---|---|
信阳 | MAE | 0.058 | 0.087 | 0.099 | 0.101 |
RMSE | 0.082 | 0.112 | 0.131 | 0.138 | |
R 2 | 0.881 | 0.787 | 0.704 | 0.664 | |
ACC | 0.962 | 0.886 | 0.835 | 0.844 | |
F 1 | 0.962 | 0.886 | 0.835 | 0.844 | |
阜阳 | MAE | 0.039 | 0.052 | 0.061 | 0.064 |
RMSE | 0.051 | 0.070 | 0.082 | 0.090 | |
R 2 | 0.924 | 0.853 | 0.806 | 0.755 | |
ACC | 0.911 | 0.911 | 0.899 | 0.857 | |
F 1 | 0.910 | 0.911 | 0.899 | 0.856 | |
霍邱 | MAE | 0.043 | 0.066 | 0.081 | 0.085 |
RMSE | 0.058 | 0.093 | 0.115 | 0.129 | |
R 2 | 0.941 | 0.839 | 0.752 | 0.615 | |
ACC | 0.937 | 0.924 | 0.899 | 0.883 | |
F 1 | 0.934 | 0.924 | 0.899 | 0.882 | |
固始 | MAE | 0.041 | 0.063 | 0.078 | 0.079 |
RMSE | 0.059 | 0.084 | 0.109 | 0.113 | |
R 2 | 0.930 | 0.853 | 0.779 | 0.730 | |
ACC | 0.924 | 0.873 | 0.835 | 0.870 | |
F 1 | 0.922 | 0.873 | 0.834 | 0.870 | |
定远 | MAE | 0.043 | 0.061 | 0.084 | 0.087 |
RMSE | 0.061 | 0.088 | 0.120 | 0.127 | |
R 2 | 0.933 | 0.851 | 0.776 | 0.671 | |
ACC | 0.950 | 0.924 | 0.873 | 0.883 | |
F 1 | 0.950 | 0.923 | 0.873 | 0.883 |