To handle the aforementioned problems, in this study, we first make an exhaustive and comprehensive review of the publicly accessible RS datasets. As a result, lots of cutting-edge and off-the-shelf deep learning methods from the machine learning community are not evaluated and compared on RS data. Due to the large variance in data collection sensors and pre-processing pipelines, it is non-trivial to directly adapt modern deep learning models to RS datasets. ![]() However, existing works usually evaluate the performance with different dataset splits, which makes it difficult to fairly and reliably compare different algorithms. For deep learning methods, the backbone networks, hyper-parameters and training tricks are influential factors that should be considered for fair performance comparison. There is a lack of an open platform for different EO tasks. Thus, it is urgent to build new RS benchmarks to enable a fair comparison of different methods. It increases the uncertainty during the evaluation of algorithms. Moreover, many RS datasets are published with no standard train/validation/test splits. However, datasets with small-scale or limited geographic coverage may bias to a specific data distribution rather than the real-world scenarios. Although several RS datasets with large volumes are published, many of the currently developed methods are still evaluated on small-scale datasets. Thus they can be more reliable for the performance validation and comparison of deep learning algorithms. Compared with small-scale datasets, large-scale ones with rich semantic annotations align better with the complex real-world scenarios. In the computer vision (CV) community, some large-scale dataset like ImageNet is usually used for the evaluation of newly developed deep learning models. ![]() There is no unified benchmark for fairly comparing remote sensing methods. As we can see, more and more, larger and larger datasets have been constructed and published during the past decade. Figure 1: Chronological overview of the volumes of existing 400 datasets. As seen from this figure, more and more, larger and larger datasets have been constructed and published during the past decade. 1, we show a chronological overview of the volumes of existing 401 datasets. In this context, recent RS datasets are constructed with larger and larger volumes. One of the key success of deep learning methods is to train model with large-scale data. To deal with large-scale data, deep learning techniques have been proven effective for many different research areas. To enable a broad range of real-world applications, a huge amount of RS data with global-coverage and high-resolution is received per day for automatic processing and analysis. With the development of Earth observation technology, more and more satellites with diverse imaging sensors are launched for different missions. The platform, dataset collections are publicly available at. The insightful results are beneficial to future research. ![]() Based on the EarthNets platform, extensive deep learning methods are evaluated on the new benchmark. EarthNets supports standard dataset libraries and cutting-edge deep learning models to bridge the gap between remote sensing and machine learning community. Furthermore, a new platform for Earth observation, termed EarthNets, is released towards a fair and consistent evaluation of deep learning methods on remote sensing data. Based on the dataset attributes, we propose to measure, rank and select datasets to build a new benchmark for model evaluation. We systemically analyze these Earth observation datasets from five aspects, including the volume, bibliometric analysis, research domains and the correlation between datasets. In this paper, for the first time, we present a comprehensive review of more than 400 publicly published datasets, including applications like, land use/cover, change/disaster monitoring, scene understanding, agriculture, climate change and weather forecasting. With an increasing number of satellites in orbit, more and more datasets with diverse sensors and research domains are published to facilitate the research of the remote sensing community. Earth observation, aiming at monitoring the state of planet Earth using remote sensing data, is critical for improving our daily lives and living environment.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |