PrediTree
收藏资源简介:
PrediTree是一个全面的开放源数据集,旨在训练和评估亚米级分辨率下的树木高度预测模型。该数据集结合了非常高的分辨率(0.5米)的激光雷达衍生的冠层高度图,与多时相和多光谱图像在空间上对齐,覆盖法国多样化的森林生态系统,总计3,141,568张图像。PrediTree通过使深度学习方法能够根据多个过去的观察来预测树木生长,填补了森林监测能力的关键空白。为了利用这个PrediTree数据集,我们提出了一个编码器-解码器框架,它需要多时相多光谱图像以及冠层高度图时间戳(目标)和每个图像获取日期之间的相对时间差异(以年为单位),以预测冠层高度。所进行的实验表明,在PrediTree数据集上训练的U-Net架构具有最高的掩码均方误差为11.78%,比次优架构ResNet-50高出约12%,并且将相同实验但在更少波段(仅红、绿、蓝)上的误差削减了约30%。这个数据集在HuggingFace上公开可用,处理和训练代码库也在GitHub上提供。
PrediTree is a comprehensive open-source dataset designed for training and evaluating tree height prediction models at sub-meter resolution. This dataset spatially aligns very high-resolution (0.5 m) LiDAR-derived canopy height maps with multi-temporal and multi-spectral images, covering diverse forest ecosystems across France, with a total of 3,141,568 images. PrediTree fills a critical gap in forest monitoring capabilities by enabling deep learning methods to predict tree growth based on multiple past observations. To leverage the PrediTree dataset, we propose an encoder-decoder framework that requires multi-temporal multi-spectral images, along with the timestamps of canopy height maps (targets) and the relative temporal differences (in years) between each image’s acquisition date, to predict canopy height. Conducted experiments show that the U-Net architecture trained on the PrediTree dataset achieves the highest mask mean squared error of 11.78%, which is approximately 12% higher than the sub-optimal ResNet-50 architecture, and reduces the error by about 30% for the same experiment using fewer bands (only red, green, blue). This dataset is publicly available on HuggingFace, and the processing and training codebase is also provided on GitHub.




