0 引言
1 数据和方法
1.1 数据来源和处理
图1 Sentinel-1影像的空间分布(红色数字表示影像序号,红色框表示影像空间位置。) Fig.1 The spatial distribution of Sentinel-1 images (The red number indicates the image serial number, and the red box indicates the image spatial position.) |
表1 Sentinel-1影像信息Tab.1 The coding list of Sentinel-1 images |
序号 | 成像日期 | 经度 | 纬度 | 卫星 | 用途 |
---|---|---|---|---|---|
1 | 2021-10-27 | 131.45°W—111.54°W | 73.83°N —78.49°N | Sentinel-1B | 测试 |
2 | 2022-01-30 | 111.25°W—95.77°W | 70.29°N —74.80°N | Sentinel-1A | 测试 |
3 | 2022-01-31 | 72.80°W—58.44°W | 68.99°N —73.47°N | Sentinel-1A | 测试 |
4 | 2022-10-31 | 134.51°W—107.33°W | 76.85°N —81.92°N | Sentinel-1A | 训练 |
5 | 2022-10-31 | 140.17°W—124.76°W | 69.97°N —74.69°N | Sentinel-1A | 训练 |
6 | 2022-11-24 | 113.12°W—93.62°W | 73.58°N —78.22°N | Sentinel-1A | 训练 |
7 | 2022-11-26 | 111.20°W—95.77°W | 70.29°N —74.80°N | Sentinel-1A | 训练 |
8 | 2022-11-27 | 96.81°W—81.42°W | 70.25°N —74.76°N | Sentinel-1A | 训练 |
9 | 2023-01-18 | 86.97°W—68.44°W | 72.78°N —77.41°N | Sentinel-1A | 测试 |
10 | 2021-12-09 | 77.70°E—99.03°E | 77.24°N —82.11°N | Sentinel-1B | 测试 |
11 | 2022-01-16 | 59.76°E—79.41°E | 73.68°N —78.33°N | Sentinel-1A | 测试 |
12 | 2022-01-27 | 112.31°E—139.43°E | 76.82°N —81.90°N | Sentinel-1A | 训练 |
13 | 2022-01-27 | 109.03°E—128.43°E | 73.55°N —78.19°N | Sentinel-1A | 训练 |
14 | 2022-11-29 | 81.86°E—101.92°E | 73.00°N —78.31°N | Sentinel-1A | 训练 |
15 | 2022-12-10 | 39.32°E—52.06°E | 66.83°N —71.26°N | Sentinel-1A | 训练 |
16 | 2023-02-04 | 87.81°E—115.34°E | 76.98°N —82.06°N | Sentinel-1A | 测试 |
17 | 2023-02-20 | 131.25°E—151.36°E | 73.03°N —78.34°N | Sentinel-1A | 训练 |
18 | 2023-03-08 | 49.27°E—64.70°E | 72.78°N —77.41°N | Sentinel-1A | 测试 |
1.2 影像预处理和假彩色合成
1.3 数据集制作
1.4 模型介绍
1.5 评价指标
2 结果和分析
2.1 切片大小对海冰分类结果的影响
图2 不同切片大小影像的海冰分类结果示例Fig.2 The examples of sea ice classification results with different slice size images |
表2 不同尺寸切片影像分类结果评价Tab.2 Evaluation for classification results of different slice size images |
区域 | 切片大小 | 准确率/% | 整体准确 率/% | Kappa 系数/% | ||
---|---|---|---|---|---|---|
海水 | 一年冰 | 多年冰 | ||||
2号 | 32×32像素 | 100.00 | 100.00 | 98.88 | 99.62 | 99.44 |
16×16像素 | 100.00 | 100.00 | 98.88 | 99.62 | 99.44 | |
8×8像素 | 99.25 | 100.00 | 99.88 | 99.71 | 99.63 | |
16号 | 32×32像素 | 100.00 | 100.00 | 97.50 | 99.17 | 98.75 |
16×16像素 | 100.00 | 99.63 | 99.38 | 99.67 | 99.50 | |
8×8像素 | 100.00 | 99.88 | 100.00 | 99.96 | 99.94 |
2.2 偏移量处理假彩色合成图像对海冰分类的影响
表3 不同假彩色合成图像的海冰分类精度评价Tab.3 Evaluation of classification accuracy of sea ice in different false color images |
区域 | 偏移量 处理情况 | 准确率/% | 整体准确 率/% | Kappa 系数/% | ||
---|---|---|---|---|---|---|
海水 | 一年冰 | 多年冰 | ||||
3号 | 未处理 | 99.88 | 99.63 | 77.63 | 92.37 | 88.56 |
处理后 | 99.25 | 100.00 | 100.00 | 99.83 | 99.75 | |
10号 | 未处理 | 99.75 | 95.75 | 96.38 | 97.29 | 95.94 |
处理后 | 99.25 | 100.00 | 100.00 | 99.75 | 99.63 |
2.3 不同深度学习模型海冰分类结果的对比分析
表4 不同深度学习模型的海冰分类精度评价Tab.4 Evaluation of classification accuracy of sea ice in different deep learning models |
区域 | 深度学习模型 | 准确率/% | 整体准确率/% | Kappa系数/% | ||
---|---|---|---|---|---|---|
海水 | 一年冰 | 多年冰 | ||||
ResNet | 99.50 | 100.00 | 93.75 | 97.75 | 96.63 | |
1号 | Vision Transformer | 98.63 | 100.00 | 99.63 | 98.12 | 97.19 |
Swin Transformer | 97.50 | 100.00 | 100.00 | 99.17 | 98.75 | |
ResNet | 100.00 | 100.00 | 94.38 | 98.12 | 97.19 | |
2号 | Vision Transformer | 96.88 | 99.88 | 100.00 | 98.92 | 98.38 |
Swin Transformer | 99.25 | 100.00 | 99.88 | 99.71 | 99.63 | |
ResNet | 100.00 | 100.00 | 98.50 | 99.50 | 99.25 | |
3号 | Vision Transformer | 95.25 | 100.00 | 98.88 | 98.42 | 97.63 |
Swin Transformer | 99.25 | 100.00 | 100.00 | 99.75 | 99.63 | |
ResNet | 99.88 | 99.88 | 97.88 | 99.17 | 98.81 | |
9号 | Vision Transformer | 98.88 | 100.00 | 100.00 | 99.62 | 99.44 |
Swin Transformer | 99.88 | 99.88 | 99.25 | 99.67 | 99.50 | |
ResNet | 100.00 | 100.00 | 97.50 | 99.62 | 99.44 | |
10号 | Vision Transformer | 97.50 | 100.00 | 100.00 | 99.17 | 98.75 |
Swin Transformer | 99.50 | 100.00 | 100.00 | 99.75 | 99.63 | |
ResNet | 100.00 | 100.00 | 99.38 | 99.79 | 99.69 | |
11号 | Vision Transformer | 99.25 | 100.00 | 100.00 | 99.83 | 99.75 |
Swin Transformer | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
ResNet | 99.50 | 100.00 | 98.00 | 99.17 | 98.75 | |
16号 | Vision Transformer | 95.13 | 100.00 | 100.00 | 98.37 | 97.56 |
Swin Transformer | 100.00 | 99.88 | 100.00 | 99.96 | 99.94 | |
ResNet | 100.00 | 100.00 | 93.88 | 97.96 | 96.94 | |
18号 | Vision Transformer | 97.13 | 100.00 | 99.88 | 99.00 | 98.50 |
Swin Transformer | 98.63 | 100.00 | 99.63 | 99.42 | 99.13 |