1 引言
2 随机森林算法及模型构建
3 资料来源
3.1 雷达产品及其指标简介
3.2 强对流天气资料
表1 随机森林模型的训练集(2008 -2017年)和测试集(2018 -2019年)中各类冰雹灾害性天气的发生次数Table 1 The occurrence frequencies of four types of hail disaster weather in the training set (2008 -2017) and the testing set (2018 -2019) used in the random forest model, respectively |
| 样本 | 冰雹灾害性天气的发生次数/次 | ||||
|---|---|---|---|---|---|
| 冰雹 | 冰雹大风 | 冰雹短强 | 冰雹大风短强 | 无强对流 | |
| 训练样本集 | 52 | 45 | 16 | 24 | 150 |
| 测试样本集 | 9 | 6 | 10 | 5 | 32 |
3.3 模型检验指标
4 结果分析
4.1 雷达产品的特征分析
图4 随机森林模型中雷达产品8个主要自变量CR (a)、 H (b)、 ET (c)、 VILD (d)、 VIL (e)、 TOP (f)、 HT (g)、 H_45 (h)在无强对流天气、 冰雹、 冰雹大风、 冰雹短强以及冰雹大风短强概率密度分布Fig.4 The probability density distribution of 8 main independent variables of the random forest model for the no convective weather, the hail weather, hail with strong wind, hail with short-time strong precipitation, and hail with strong and short-time strong precipitation, the radar products of CR (a), VIL (b), ET (c), HT (d), H (e), TOP (f), VILD (g) and H_45 (h) |
4.2 随机森林模型在训练期的判识结果
表2 随机模型训练期(2008 -2017年)的分类识别结果Table 2 The identification results of the random forest model for the training set during 2008 -2017 |
| 实况观测 | 随机森林模型结果 | ||||
|---|---|---|---|---|---|
| 冰雹 | 冰雹大风 | 冰雹短强 | 冰雹大风短强 | 无强对流 | |
| 冰雹 | 48 | 2 | 0 | 2 | 0 |
| 冰雹大风 | 2 | 42 | 0 | 1 | 0 |
| 冰雹短强 | 1 | 0 | 14 | 1 | 0 |
| 冰雹大风短强 | 0 | 1 | 2 | 21 | 0 |
| 无强对流 | 0 | 0 | 0 | 0 | 150 |
表3 随机模型训练期(2008 -2017年)的分类识别结果评分Table 3 The POD, FAR and CSI scores for the classified identification of random forest model in the training set (from 2008 to 2017) |
| 天气类型 | POD | FAR | CSI |
|---|---|---|---|
| 冰雹 | 92.3% | 5.9% | 87.3% |
| 冰雹大风 | 93.3% | 6.7% | 87.5% |
| 冰雹短强 | 87.5% | 12.5% | 77.8% |
| 冰雹大风短强 | 87.5% | 19.1% | 75.0% |
| 平均值 | 90.2% | 11.1% | 81.9% |
4.3 随机森林模型的独立样本测试
表4 随机森林模型测试集(2018 -2019年)的判识结果Table 4 The identification results of the random forest model for the testing set from 2018 to 2019 |
| 实况观测 | 随机森林模型判识结果 | ||||
|---|---|---|---|---|---|
| 冰雹 | 冰雹大风 | 冰雹短强 | 冰雹大风短强 | 无强对流 | |
| 冰雹 | 7 | 1 | 0 | 0 | 1 |
| 冰雹大风 | 1 | 5 | 0 | 0 | 0 |
| 冰雹短强 | 1 | 0 | 7 | 2 | 0 |
| 冰雹大风短强 | 0 | 1 | 1 | 3 | 0 |
| 无强对流 | 3 | 2 | 0 | 0 | 27 |
表5 随机森林模型测试集(2018 -2019年)的判识结果评分Table 5 The POD, FAR and CSI scores for the classified identification of random forest model in the testing set (from 2018 to 2019) |
| 天气类型 | POD | FAR | CSI |
|---|---|---|---|
| 冰雹 | 77.8% | 41.7% | 50.0% |
| 冰雹大风 | 83.3% | 44.4% | 50.0% |
| 冰雹短强 | 70.0% | 12.5% | 63.6% |
| 冰雹大风短强 | 60.0% | 40.0% | 42.9% |
| 平均值 | 72.8% | 34.7% | 51.6% |
图6 2018年6月6日(a, b)和10日(c, d)强对流天气实况和随机森林模型的分类识别结果三角形、 圆点、 星形分别代表短时强降水、 冰雹和大风; 色斑为时段内的雷达产品CR的值 Fig.6 Example of severe convective weather for observations and the classified results of random forest model on 6 (a, b) and 10 (c, d) June 2018.Triangle, round point and square mean the occurrence of short-time strong precipitation, hail and strong wind, respectively; Colorful shadow is the radar product of CR |
4.4 随机森林模型的预报效果评估
表6 随机森林模型的预报效果评分Table 6 The prediction scores of random forest model |
| 天气类型 | POD | FAR | CSI |
|---|---|---|---|
| 冰雹 | 78.8% | 30.7% | 61.6% |
| 冰雹大风 | 84.3% | 24.4% | 65.0% |
| 冰雹短强 | 70.8% | 12.5% | 63.6% |
| 冰雹大风短强 | 65.3% | 30.0% | 52.8% |
| 平均值 | 74.8% | 24.4% | 60.8% |
图7 2019年4月15日(a, b)和6月12日(c, d)强对流天气实况和随机森林模型预报结果对比三角形、 圆点和星形分别代表短时强降水、 冰雹和大风; 色斑为时段内的CR Fig.7 Example of severe convective weather for observations and the forecasting results of random forest model on 15 April (a, b) and 12 June (c, d) 2019.Triangle, round point and square mean the occurrence of short-time strong precipitation, hail and strong wind, respectively, and colorful shadow is the radar product of CR |