Conference Agenda

This study investigated the use of C- and X-band SAR data for estimation of forest changes (height, biomass and biomass change) in a boreal forest in Sweden. Field plot data (10 m plots) from 2016 and 2021, and lidar data were used as references. Plots with substantial decreases of biomass (due to clear-cuts and thinning) could be detected using Radarsat-2 normalized backscatter images (C-band) while the use of interferometry (InSAR) of TanDEM-X images allowed both accurate biomass and biomass change estimation, and mapping of smaller forest height changes (increase). We conclude that both C- and X-band SAR (Radarsat-2 and TanDEM-X) are useful for estimation of forest decline, while the use of TanDEM-X InSAR provides added value in terms of height information and therefore more accurate estimates of biomass and biomass change.

With the aggravation of global warming, the outbreaks of forest pests happen more frequently and damage huge amounts of forests. Detecting and removing trunk-boring infestations from the forest at an early stage (green attack) is important to avoid spreading. By using remote sensing techniques, forest mortality can be efficiently detected and mapped; however, achieving early detection of the infestation is still challenging because the spectral changes are subtle.

This study assessed the detectability of the green attacks by the European spruce bark beetle (Ips typographus, L.) using multispectral drone images. The scientific questions and objectives are (1) testing whether the infestations showed detectable vulnerability before attacks, (2) quantifying the detectability of the attacked trees with different duration of infestations, (3) testing the detection performance using a single vegetation index (VI) in comparison with machine learning models with multiple variables, and testing their performance when applied on untrained areas, and (4) testing the performance of the MR_DSWI2 index we proposed in a previous study.

The study used multispectral drone images covering 24 plots from 6 forest stands in southern Sweden, acquired in May (before attacks), June (green attack), August (green and yellow attack), and October 2021 (red attack). Weekly field inventory was conducted on 997 spruce trees and the starting weeks of the infestations were recorded for 208 attacked trees. Drone images of individual-tree crowns were segmented using marker-controlled watershed segmentation, and 10 VIs [5] were calculated for every single tree. Trees with the same duration of infestation were grouped for the analysis. Random Forest Classification (RF) and linear discriminant analysis (LDA) models were built using (1) all bands, (2) all VIs, and (3) the four bands used for MR_DSWI2. Three LDA models were also built using MR_DSWI2, NDRE2, and NGRDI indices, respectively. To test the potential overfitting and test the applicability of the models on mapping untrained areas, two ways of separating training and testing data were used and compared. Method A trained the model using trees from 5 stands and tested on the trees from the remaining stand. This design tested the performance of the models in untrained areas, and overfitted models would yield low accuracy. Method B did not separate trees from different stands, but randomly assigned 90% of all trees for training and the other 10% for testing. Method B is commonly used for separating training and testing dataset, but it could not show the potential performance when applied to untrained areas. When comparing the results from Method A and B, a potential overfitting could be exhibited. The classification accuracy was presented using the Kappa Coefficient.

The results and conclusions are: (1) When using models with more dimensions and higher complexity (i.e., the RF models), the detection had high accuracy in the trained areas but low accuracy in untrained areas. Overfitting was more prominent with those models, and Method B, i.e., testing models with data from the same areas as the training data, could not indicate the applicability outside the study area. Thus, testing the model performance using Method A was proposed and results were further discussed. (2) When testing on untrained areas (Method A, Figure 1), no model successfully classified healthy and attacked trees before the attacks. We conclude that the attacked trees showed no vulnerability before the attacks. (3) When testing on untrained areas (Method A, Figure 1), no model successfully separated healthy and attacked trees when infested for less than 5 weeks. The green attacks during 1-5 weeks of infestation were almost undetectable. During 10–13 weeks of infestation, the detectability increased, with 0.67–0.75 of the median kappa coefficients using the best classification model (LDA by MR_DWSI2, Figure 1). Trees infested for 19–22 weeks (red attack) showed high detectability. (4) Among all classification models, using LDA on the MR_DWSI2 index achieved the highest and most stable accuracy at various infestation stages, followed by LDA using the four bands and LDA on NGRDI.

【Objective】Using the five bands (P/L/S/C/X) of quad-pol SAR data acquired by the High-resolution Airborne System, this study analyzed the response law and sensitivity of different band signals to forest stem volume. Moreover, the capability of estimating stem volume based on both single-frequency and multi-frequency PolSAR data was evaluated. 【Method】Combined with field measurements and airborne LiDAR data, forest stem volume was scaled-up to the entire study area, and a total of 196 samples were obtained through stratified sampling. Based on the samples, the forest stem volume estimation capability of multi-band polarimetric SAR was evaluated. The Water Cloud Model was utilized to analyze the response law of SAR backscattered intensity to stem volume across different bands, and the dynamic range and saturation point were quantified. Furthermore, multiple polarimetric decomposition components were extracted and their sensitivity to stem volume was analyzed using correlation coefficients. On this basis, random forest and support vector regression algorithms were used to perform feature selection and regression modeling. 【Result】The Water Cloud Model analysis revealed that the dynamic range of the longer wavelength (P/L) is higher compared to the shorter wavelength (S/C/X). In particular, the saturation point for the P band exceeds 160 m³/ha, whereas it does not surpass 100 m³/ha for the other bands.The correlation analysis results indicated that the correlation between the P band, L/S bands, and C/X bands and stem volume decreases in order, with values above 0.6, between 0.3-0.4, and below 0.3, respectively. When estimating stem volume using a single band, the accuracy of the P band was 73.79%, while other bands did not exceed 60%. When using multi-band joint estimation, the combination of L/S and P bands improved the estimation accuracy by approximately 2% compared to using the P band alone. The contribution of adding the C/X band to the accuracy improvement was minimal. The best estimation result was obtained by combining all the bands, achieving an accuracy of 77.25%. 【Conclusion】Taking into account factors such as signal dynamic range, saturation point and correlation, the P band is the most sensitive to forest stem volume, which is significantly better than the other bands. The L/S bands are second in sensitivity, while the C/X bands are the least sensitive. When estimating forest stem volume using PolSAR data, the P band should be the first choice. Additionally, when using multi-band joint estimation, the combination of P and L/S bands should be preferred.

Cold-temperate Forest (CTF) is a vital carbon sink and source for the world economy but has been significantly disturbed by intensified human activities and climate change such as deforestation, fires, pests, diseases, etc., resulting in a significant degradation trend, which has an impact on the regional and global carbon budget process and its assessment. However, the pattern of forest disturbance in CTF in northern China is not well known. In this paper, Genhe forest area, a typical CTF region located in Inner Mongolia Autonomous Region, Northeast China (about 2.001×10⁴ km²), was selected as the study area. Based on the Landsat historical archived data on the Google Earth Engine (GEE) platform, we used the Continuous Change Detection and Classification (CCDC) algorithm incorporating spectral indices and seasonal characteristics to detect forest disturbances in nearly 30 years. First, we created six (2 seasons × 3 indices) interannual time-series seasonal vegetation index datasets to map the forest coverage using the between-class variance algorithm (OTSU). Second, we improved the CCDC algorithm incorporating with vegetation index and seasonal characteristics to extract the disturbance information. Finally, we evaluated how the disruption relates to the climate and human activity. The results showed that the disturbance map produced by using summer (June-August) imagery and the EVI vegetation index had the highest overall accuracy of 84.15%. Forests have been disturbed to the extent of 12.65% (2137.31km²) over the last 30 years, and the disturbed area generally showed a trend toward reduction, especially after commercial logging activities were banned in 2015. But there was an unusual increase in the disturbed area in 2002 and 2003 due to large fires. Monitoring of potential widespread forest disturbance due to extreme drought and fire events in the context of climate change should be strengthened in the future, and preventive and salvage measures should be taken in a timely manner. Our results demonstrates that CTF disturbance can be robustly mapped using CCDC algorithm based on Landsat time-series seasonal imagery in areas with complex meteorological conditions and spatial heterogeneity, which is essential for understanding forest change processes.

The multi-temporal polarimetric SAR data provides the difference of scattering characteristics in time dimension for scene observation, hence it could reflect the time-variant characteristics of the same scene. Based on this advantage, classification is one of the important applications of multi-temporal polarimetric SAR data. However, the features of time and polarization dimension used for classification basically are from the data at each certain time, which lack the interpretation of the scattering variant characteristics between multi-temporal data. To solve the problem, based on the specific data representation models for multi temporal polarimetric SAR data, this paper extracts new time-variant scattering features, including the change type as well as the change direction of scattering, which the previous static temporal/polarimetric features cannot provide.

Time series Radarsat-2 data is used for experiments. It includes 8 Fully PolSAR images from April 14,2009 to September 29,2009. The data interval of each scene is 24 days. The image size is 5300*3100 pixels. It contains 22 categories. Based on the difference model T_m=T_i-r_mT_j（T_i,T_j are polarization coherence matrices of two times polarimetric SAR data, and r_m is the smallest eigenvalue of T_i^-1T_j, it can make sure that T_m is always a positive semidefinite Hermitian matrix）, we extract a series of time-variant scattering features, using the components of H/A/α decomposition and Freeman decomposition to classify. Through analysis the classification results of transformer classifier, compared with traditional static features, using the proposed time-variant scattering features to classify can effectively improve the classification accuracy.

As a kind of active microwave remote sensing, SAR (Synthetic Aperture Radar) is very suitable for the resolution of different features with its all-weather and all-day characteristics. In order to improve the resolution of different features and make in-depth analysis and research on different feature types, it is necessary to use the multi-band and multi-polarization information of SAR. The transmission characteristics and the backscattering characteristics of target echoes are different for SAR of different bands, and the fusion of SAR images of different bands can better integrate the information of SAR images of different bands.

In this paper, the experimental data were selected from the airborne data acquired by two Xinzhou 60 remote sensing aircraft modified by the Air and Space Academy of Chinese Academy of Sciences under the support of the Chinese Aeronautic Remote Sensing System (CARSS) construction project. Fully PolSAR data including C and S bands contains five types such as paddy fields, forested lands, dry lands, artificial buildings and water. And it is classified using scattering features such as H/A/α and Freeman decomposition components. In order to make full use of the advantages of multi-band, a multi-band fully PolSAR image fusion method based on wavelet transform is also proposed in the paper, which takes advantage of the wavelet transform for multi-resolution fusion and combines the variation of scattering features of different feature types by different bands, and is applied to airborne C-band and S-band SAR images to realize the classification of different types of features in the same area, which can effectively improve the classification effect and increase the classification accuracy.

Polarized synthetic aperture radar (PolSAR) can obtain rich feature information by receiving electromagnetic waves from different features at different polarization combinations, and is widely used in fields such as feature classification. The coherence of multi-temporal polarization SAR data is a useful supplement to polarization SAR data, which contains information that is not available in single phase polarization data. This paper aims to introduce temporal coherence analysis into dual polarization data information and coherently process SAR images acquired at different times in the same region, so as to effectively combine the information of both time dimension and polarization dimension and improve the accuracy of crop classification of Sentinel data. In order to fully extract the changes of feature in the time dimension, this paper obtain the distinction and connection of each category in temporal features by using the multi-temporal feature information of SBAS. Based on multi-temporal polarization SAR coherence optimization, Sentinel-1 data is used for experimental analysis and verification. Data were collected from the time series dual-polarization data of Yucatan Lake area from April to September 2019. The effects of different polarization states on the characteristics of crop species is explored and the classification effects of the optimized values of coherence features is analyzed.

With the BIOMASS mission, developed by the European Space Agency, P-band polarimetric interferometric synthetic aperture radar (PolInSAR) is expected to provide a fresh perspective on the estimation of above-ground biomass in global forests. However, in sparse boreal forest areas, the strong penetration of P-band interferometric SAR (InSAR) can lead to a loss of forest canopy information, resulting in a biased estimation of forest AGB. In contrast, X-band InSAR signals are sensitive to forest canopy information. By combining the two type approaches, forest parameters such as height, density, and AGB can be effectively extracted. Furthermore, there are a number of X-band InSAR satellite systems which is expected to be launched or already in operation, such as Tandem-X. The data from those systems has great potential to collaborate with BIOMASS data to estimate forest AGB. In summary, a novel forest AGB estimation algorithm based on dual-frequency InSAR data has been proposed for accurate estimation of above-ground biomass in boreal forests. This approach utilizes the difference in penetration ability of P-band and X-band InSAR in forest areas to extract forest height without bias. Based on this, a multi-dimensional feature set has been constructed, including direct information such as forest height and density, as well as indirect information such as backscatter intensity, polarization features, and image texture, to achieve the estimation of forest AGB. To verify the method, experiments were conducted based on airborne P-band PolInSAR data and spaceborne X-band InSAR data.

Synthetic aperture radar tomography (TomoSAR) is a three-dimensional imaging technique developed on the basis of multi-baseline SAR interferometry (MB-InSAR). In forest scenarios, TomoSAR can achieve vertical characterization of forest scatterers, which is an effective method for extracting structural parameters such as forest height and above-ground biomass (AGB). However, due to factors such as baseline errors and changing atmospheric conditions, phase errors between MB-InSAR data cause severe sidelobe effects or even complete defocusing in tomography imaging.

In the study, we compared the calibration methods based on polynomial fitting (PF) and entropy minimization (EM), as well as proposed a novel approach combine the two methods. All three methods are tested based on airborne P-band MB-InSAR data, and the accuracy of the forest height extracted based on proposed method is verified based on airborne light detection and ranging (LiDAR) data. he results indicate that the proposed approach outperforms the other two methods in term of tomography imaging and can accurately determine forest height with an accuracy of over 80%.

Grassland is an important component of terrestrial ecosystems, but due to human activities and natural changes, the productivity and ecological service capacity of grassland ecosystems have declined. Ecological environmental problems such as land desertification and grassland degradation have become hot topics of global concern. Therefore, timely and accurate monitoring of changes in grassland type distribution, vegetation utilization, and intensity is of irreplaceable importance for protecting the ecological environment. High- and medium-resolution optical imagery is the most commonly used data source for grassland remote sensing monitoring. However, due to limitations in data acquisition capabilities, it is not possible to obtain time-continuous data using a single data source, which affects the precise monitoring of grassland distribution, utilization, and intensity. With the increasing availability of different high- and medium-resolution remote sensing data, the fusion of multiple data sources to generate high spatiotemporal resolution data for grassland monitoring has been widely applied. However, there is currently limited research on grassland monitoring that considers the fusion of China's high-resolution data with other medium- to high-resolution data without introducing low spatial resolution data information. To address this issue, this study aims to use a combination of China's satellite products and other high- to medium-resolution optical remote sensing data to generate daily high spatiotemporal resolution surface reflectance data for grassland monitoring. This study was conducted in a 50×50 km study area in Hulunbuir grassland. First, based on the coordinated satellite products Landsat, Sentinel-2 (HLS) data, and GF-6 WFV image data, the spectral reflectance conversion equation between the two was constructed using the least squares method to transform the reflectance of GF-6 WFV. Second, based on the acquired real image data set, the local linear interpolation method was used to fill in the missing images to generate a daily data set. The Savitzky-Golay spatial filtering algorithm was used to smooth and denoise the time series data set to construct the daily high spatiotemporal resolution surface reflectance data set (HLSG). Finally, based on the time series vegetation index information, the grassland utilization intensity index was proposed. The study showed that time series reconstruction based on GF-6 WFV and HLS data can solve the problems of data missing, quality, and accuracy, thereby improving the reliability and accuracy of medium spatial resolution time series data. Moreover, the grassland utilization intensity estimation method based on the daily NDVI time series data set can well reflect the differences in grassland utilization intensity and has important significance for monitoring the utilization status of grasslands.

Vegetation is an important material foundation for the ecological function of grassland ecosystem, and the long-term changes in vegetation status can provide important information reference for the study on evolution laws of various ecosystems. Therefore, based on the normalized vegetation index product (NDVI, MOD13A2) of long time series (2000-2020) MODIS data, an improved directional pixel binary model was used to construct the annual vegetation coverage dataset of the Beijing-Tianjin Sandstorm Source Area (BTSSA) from 2001 to 2021. Research methods such as Sen-Mann Kendall time series trend significance test and spatial statistical analysis were used to analyze the spatiotemporal changes in vegetation coverage in the region over the past 20 years. The results showed that in the past 20 years, 50.94% of the study area had significantly decreased vegetation coverage, while less than 2% had significantly increased vegetation coverage. This indicates that due to urban development and climate change, the vegetation situation in the BTSSA has caused some damage to a certain extent. These research results, combined with climate and underlying surface change data, can provide important information support for subsequent research on grassland ecosystem.