Search

The self-attention mechanism has been empirically shown its effectiveness in a wide range of computer vision applications. However, it is usually criticized for the expensive computation cost. Although some revised methods are proposed in the recent past, they are not maturely applicable to remote sensing scene (RSS) images. To address this problem, in this article, we propose a simple yet effective context acquisition module, named thrifty attention, which can capture the long-range dependence efficiently and effectively. Moreover, a recurrent version for thrifty attention, termed recurrent thrifty attention (RTA), is further proposed to take the long-range multihop communications in space–time for RSS images. RTA is a general global contextual information acquisition module that can be used in any hierarchy of deep convolutional neural networks. To demonstrate its superiority, we deploy it to the classical ResNet and establish our proposed RTA Network (RTANet). Extensive experiments are carried out on two levels of the RSS recognition tasks, i.e., the image-level RSS classification and the instance-level RSS object detection. Compared with the standard self-attention mechanism, RTA can reduce at most 0.43 M model parameters while increasing a slight of model floating-point operations per second (FLOPs). Furthermore, results on RSS classification and object detection further verify the accuracy superiority of RTANet. [ABSTRACT FROM AUTHOR]

Background: Fast and accurate forest aboveground biomass (AGB) estimation and mapping is the basic work of forest management and ecosystem dynamic investigation, which is of great significance to evaluate forest quality, resource assessment, and carbon cycle and management. The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2), as one of the latest launched spaceborne light detection and ranging (LiDAR) sensors, can penetrate the forest canopy and has the potential to obtain accurate forest vertical structure parameters on a large scale. However, the along-track segments of canopy height provided by ICESat-2 cannot be used to obtain comprehensive AGB spatial distribution. To make up for the deficiency of spaceborne LiDAR, the Sentinel-2 images provided by google earth engine (GEE) were used as the medium to integrate with ICESat-2 for continuous AGB mapping in our study. Ensemble learning can summarize the advantages of estimation models and achieve better estimation results. A stacking algorithm consisting of four non-parametric base models which are the backpropagation (BP) neural network, k-nearest neighbor (kNN), support vector machine (SVM), and random forest (RF) was proposed for AGB modeling and estimating in Saihanba forest farm, northern China. Results: The results show that stacking achieved the best AGB estimation accuracy among the models, with an R2 of 0.71 and a root mean square error (RMSE) of 45.67 Mg/ha. The stacking resulted in the lowest estimation error with the decreases of RMSE by 22.6%, 27.7%, 23.4%, and 19.0% compared with those from the BP, kNN, SVM, and RF, respectively. Conclusion: Compared with using Sentinel-2 alone, the estimation errors of all models have been significantly reduced after adding the LiDAR variables of ICESat-2 in AGB estimation. The research demonstrated that ICESat-2 has the potential to improve the accuracy of AGB estimation and provides a reference for dynamic forest resources management and monitoring. [ABSTRACT FROM AUTHOR]