High spatial resolution commercial satellite data provide new opportunities for terrestrial monitoring. The recent availability of near-daily 3 m observations provided by the PlanetScope constellation enables mapping of small and spatially fragmented burns that are not detected at coarser spatial resolution. This study demonstrates, for the first time, the potential for automated PlanetScope 3 m burned area mapping. The PlanetScope sensors have no onboard calibration or short-wave infrared bands, and have variable overpass times, making them challenging to use for large area, automated, burned area mapping. To help overcome these issues, a U-Net deep learning algorithm was developed to classify burned areas from two-date Planetscope 3 m image pairs acquired at the same location. The deep learning approach, unlike conventional burned area mapping algorithms, is applied to image spatial subsets and not to single pixels and so incorporates spatial as well as spectral information. Deep learning requires large amounts of training data. Consequently, transfer learning was undertaken using pre-existing Landsat-8 derived burned area reference data to train the U-Net that was then refined with a smaller set of PlanetScope training data. Results across Africa considering 659 PlanetScope radiometrically normalized image pairs sensed one day apart in 2019 are presented. The U-Net was first trained with different numbers of randomly selected 256 × 256 30 m pixel patches extracted from 92 pre-existing Landsat-8 burned area reference data sets defined for 2014 and 2015. The U-Net trained with 300,000 Landsat patches provided about 13% 30 m burn omission and commission errors with respect to 65,000 independent 30 m evaluation patches. The U-Net was then refined by training on 5,000 256 × 256 3 m patches extracted from independently interpreted PlanetScope burned area reference data. Qualitatively, the refined U-Net was able to more precisely delineate 3 m burn boundaries, including the interiors of unburned areas, and better classify “faint” burned areas indicative of low combustion completeness and/or sparse burns. The refined U-Net 3 m classification accuracy was assessed with respect to 20 independently interpreted PlanetScope burned area reference data sets, composed of 339.4 million 3 m pixels, with low 12.29% commission and 12.09% omission errors. The dependency of the U-Net classification accuracy on the burned area proportion within 3 m pixel 256 × 256 patches was also examined, and patches <6.5% burned were less accurately classified. A regression analysis between the proportion of 30 m grid cells classified as burned against the proportion labelled as burned in the 3 m reference maps showed high agreement (r² = 0.91, slope = 0.93, intercept <0.001), indicating that the commission and omission errors largely compensate at 30 m resolution.

高空间分辨率商业卫星数据为地面监测提供了新的机会。最近，PlanetScope 星座提供了近乎每日 3 m 的观测数据，从而能够绘制在较粗空间分辨率下无法检测到的小型和空间碎片烧伤的地图。这项研究首次展示了自动化 PlanetScope 3 m 烧毁区域测绘的潜力。PlanetScope 传感器没有板载校准或短波红外波段，并且具有可变的通过时间，使其难以用于大面积、自动化、烧毁区域测绘。为了帮助克服这些问题，开发了一种 U-Net 深度学习算法，用于从在同一位置获取的两个日期的 Planetscope 3 m 图像对中对烧毁区域进行分类。深度学习方法，与传统的烧毁区域映射算法不同，它应用于图像空间子集而不是单个像素，因此结合了空间和光谱信息。深度学习需要大量的训练数据。因此，迁移学习是使用预先存在的 Landsat-8 派生的烧毁区域参考数据来训练 U-Net，然后使用较小的 PlanetScope 训练数据集进行改进。介绍了考虑到 2019 年相隔一天感知到的 659 个 PlanetScope 辐射归一化图像对的整个非洲的结果。U-Net 首先使用从 2014 年和 2015 年定义的 92 个预先存在的 Landsat-8 烧毁区域参考数据集中提取的不同数量的随机选择的 256 × 256 30 m 像素块进行训练。U-Net 使用 300 个、相对于 65,000 个独立的 30 m 评估斑块，000 个 Landsat 斑块提供了约 13% 的 30 m 燃烧遗漏和调试错误。然后通过对从独立解释的 PlanetScope 烧毁区域参考数据中提取的 5,000 个 256 × 256 3 m 块进行训练，对 U-Net 进行了改进。定性地，改进的 U-Net 能够更精确地描绘 3 m 燃烧边界，包括未燃烧区域的内部，并更好地分类指示低燃烧完整性和/或稀疏燃烧的“微弱”燃烧区域。改进的 U-Net 3 m 分类准确度针对 20 个独立解释的 PlanetScope 烧毁区域参考数据集进行了评估，这些数据集由 3.394 亿个 3 m 像素组成，具有低 12.29% 的佣金和 12.09% 的遗漏错误。还检查了 U-Net 分类精度对 3 m 像素 256 × 256 块内烧毁区域比例的依赖性，小于 6.5% 烧毁的块分类准确度较低。归类为烧毁的 30 m 网格单元的比例与 3 m 参考地图中标记为烧毁的比例之间的回归分析显示高度一致（r²  = 0.91，斜率 = 0.93，截距 <0.001），表明在 30 m 分辨率下，佣金和遗漏误差在很大程度上得到了补偿。