BigPlantSens - Assessing the Synergies of Big Data and Deep Learning for the Remote Sensing of Plant Species

包括研究、自然保护和经济活动（如林业、农业或生态系统服务评估）在内的各种任务都需要关于植物物种地理分布的准确信息。由于新的非常高空间分辨率卫星任务和无人机（UAV），越来越多的地球观测数据的可用性揭示了植被格局的高时空细节。因此，需要有效的方法来利用这一日益增长的信息来源进行植被分析。在植被遥感领域，卷积神经网络（CNN）等深度学习方法目前正在为模式和目标识别带来革命性的可能性。因此，预计随着高分辨率传感器技术的进步，CNN将扩大我们确定空间显性植被模式的能力。然而，CNN通常需要大量的参考观测来学习关键的图像特征。大数据方法可以提供训练CNN模型所需的参考观测值。各种倡议（例如，CLEF； GBIF, Pl@ntNet）提供了大量的植物物种标记图像数据，即附有物种名称的照片。由于开放数据领域的不断努力，这些图像数据集可以免费访问，并且还在继续增长。然而，目前尚不清楚用这些图像数据集训练的CNN模型在空间分辨率、质量和观看几何形状方面是否直接适用于高分辨率地球观测数据。因此，在拟议的项目中，我们的目标是评估大数据与高空间分辨率地球观测数据在全自动植被制图中的协同作用。所提出的方法使用带有物种名称标记的免费图像的大数据来训练CNN模型。然后将训练好的模型应用于高分辨率的地球观测数据，以揭示目标物种的空间分布。因此，我们试图确定用于训练的图像的哪些特征会影响映射精度（例如，采集几何形状，图像质量），并且我们将开发一种在训练前根据这些特征过滤图像数据集的算法。具体而言，我们的研究问题是：1)利用所提出的结合深度学习和高分辨率遥感数据的大数据方法，如何准确地识别不同植物物种的空间分布？2)决定基于大数据的图像数据集对CNN训练价值的关键因素是什么，这些因素是否可以通过深度学习有效地过滤？3)对地观测数据的空间分辨率如何限制植物物种识别？

Various tasks - including research, nature conservation, and economic activities such as forestry, agriculture, or ecosystem service assessments - require accurate information on the geographical distribution of plant species. Due to novel very high spatial resolution satellite missions and Unmanned Aerial Vehicles (UAV), there is a growing availability of Earth observation data revealing both high spatial and temporal detail on vegetation patterns. Consequently, efficient methods are needed to harness this growing source of information for vegetation analysis.In the field of remote sensing of vegetation, Deep Learning methods such as Convolutional Neural Networks (CNN) are currently revolutionizing possibilities for pattern and object recognition. Thus, it is expected that, in tandem with advances in high-resolution sensor technology, CNN will enlarge our capability to determine spatially explicit vegetation patterns. However, CNN commonly require ample reference observations to learn the pivotal image features. A big data approach may provide these reference observations required for training the CNN models. Various initiatives (e.g., CLEF; GBIF, Pl@ntNet) provide a vast amount of labelled image data on plant species, i.e., photographs together with species names. As a result of the constant efforts in the area of Open Data, such image datasets are freely accessible and continue to grow. However, it remains unclear if CNN models trained with such image datasets are directly applicable to very-high-resolution Earth observation data in terms of their spatial resolution, quality, and viewing geometries. Accordingly, in the proposed project, we aim to assess the synergies of big data with high spatial resolution Earth observation data for fully automated vegetation mapping. The proposed approach uses big data in terms of freely available imagery tagged with species names to train CNN models. The trained models are then applied to high-resolution Earth observation data to reveal the spatial distribution of the target species. Thereby, we seek to identify which characteristics of the images used for training affect mapping accuracy (e.g., acquisition geometry, image quality), and we will develop an algorithm for filtering the image datasets according to these characteristics before training. Specifically, our research questions are:1) How accurately can the spatial distribution of different plant species be identified using the proposed big data approach combined with deep learning and very-high-resolution remote sensing data?2) What are the critical factors determining the value of Big Data-based image datasets for CNN training, and can these be efficiently filtered using deep learning?3) How does the spatial resolution of the Earth observation data limit the plant species identification?