Core tip: Recently, the team of Professor Meng Ran from the School of Resources and Environment of our university has cooperated with the University of Maryland, Southern Methodist University, Ant Group, the University of Hong Kong, the University of Singapore, Sun Yat-sen University, Central China Normal University and other units, and has made important progress in remote sensing intelligent monitoring of the ecological environment.
Recently, the team of Meng Ran, a teacher from the School of Resources and Environment, has cooperated with the University of Maryland, Southern Methodist University, Ant Group, the University of Hong Kong, the University of Singapore, Sun Yat-sen University, Central China Normal University and other units to make important progress in the intelligent monitoring of the ecological environment through remote sensing.
Figure 1. Transformer4SITS network diagram
Fig. 2.Time series characteristics and attention distribution information of tree species remote sensing, with error bars representing standard deviation in statistics
Plantations have expanded rapidly over the past few decades and have gradually become an important part of the global ecosystem, exerting significant ecological benefits. It is of great significance to monitor the distribution of tree species in plantations based on remote sensing data for the management and protection of plantations. However, a large number of input features, complex time series and structural information in the time series of remote sensing images greatly increase the difficulty of dimensionality and feature extraction of time-spectral features, which poses a challenge to most general-purpose classifiers that lack the ability to learn time series features. To solve this problem, Meng Ran's team developed an adaptive key spectral time feature extraction method for remote sensing image time series (Transformer4SITS, Fig. 1) based on the deep learning Transformer network, which can effectively adaptively extract the time-spectral features of tree species with high separability and effectively improve the classification accuracy of tree species in a large area (Fig. 2). The research results were published in ISPRS Journal of Photogrammetry and Remote as "A spectral-temporal constrained deep learning method for tree species mapping of plantation forests using time series Sentinel-2 imagery". Published by Sensing.
Figure 3. Schematic diagram of a multi-level spatiotemporal optimization method based on S1 satellite data
Figure 4. The seasonality, size and rate of formation of deforestation patches are expressed as averages within a hexagonal grid of 5 square kilometres.
Accurate dynamic monitoring of deforestation, such as timing of occurrence, rate of formation, and scale attributes, can help better understand the impacts of deforestation in forest ecosystems and support sustainable forest management. Previous deforestation monitoring studies have focused on interannual dynamic monitoring or near-real-time deforestation discrimination, and there is a lack of methods to accurately map deforestation patches at higher temporal resolution to better reveal the fine-scale temporal dynamics of deforestation in frontlands.
Meng Ran's team studied the problems of monthly deforestation mapping of cloud-and-rain Sentinel-1 (S1) satellite data (such as the existence of speckle noise and the limitation of changes in land surface moisture) under cloudy and rainy conditions, and proposed an optimization method based on deep learning to integrate spatiotemporal background information (Fig. 3), which improved the accuracy of tropical deforestation boundary delineation based on S1 and could more accurately map the spatiotemporal distribution of deforestation. Further analysis of the spatial and temporal distribution of deforestation shows that deforestation in different parts of the tropics has seasonal characteristics and differences (Figure 4), which are essential for assessing ecological consequences, accurately accounting for carbon sequestration in tropical forests, and formulating scientific management measures. The research results were published in Remote Sensing of "Integration of deep learning algorithms with a Bayesian method for improved characterization of tropical deforestation frontiers using Sentinel-1 SAR imagery". Environment.
The two research results were signed by Huazhong Agricultural University, and Sun Rui, a doctoral student, and Huang Zehua, a postdoctoral student, were the first authors of the paper, and Meng Ran and Zhao Feng, a teacher from Central China Normal University, were the corresponding authors of the paper. His research was supported by the National Natural Science Foundation of China, the National Key R&D Program of China, the Key R&D Program of Heilongjiang Province, and the Interdisciplinary Research Institute of Huazhong Agricultural University.
Scholars of our university have made a series of progress in remote sensing intelligent monitoring of ecological environment_Scientific Research_News_Nanhu News (hzau.edu.cn)