Conference Agenda

The multi-temporal PolSAR 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. Specific multi-temporal PolSAR data representation models can initially combine time and polarization dimensions to characterize scattering changes, but the properties of such data representation models are not perfect, and the number and characteristics of extracted features are insufficient. In this paper, a power data representation model is constructed, which can maintain a complete matrix property and fully represent the changes of various scattering mechanisms, providing an effective matrix basis for the scattering changes of multi-temporal PolSAR data. Moreover, time-variant scattering features are extracted based on the power model, which can provide the types and directions of scattering changes. They are used with static features of each time for classification.

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 21 categories. Compared with the current multiplication model and the difference data representation model, the properties of the features extracted based on the power model are more comprehensive, and the classification accuracy can be effectively improved by using the extracted time-variant scattering features.

Synthetic aperture radar interferometry (InSAR) is an effective tool for investigating wide-area active landslides. However, ground deformation information derived from InSAR is susceptible to atmospheric delay due to its uneven distribution, especially within steep terrains. In general, atmospheric delay correction methods can be broadly divided into those based on external auxiliary information and those based on traditional empirical models. Since external auxiliary information is hampered by low spatial-temporal resolution, and conventional empirical phase-elevation models have limitations in window segmentation and model selection, neither can exactly reflect the heterogeneity of the atmosphere in space. In this paper, we propose an improved atmospheric delay correction method based on the multivariable moving-window empirical model (MMEM) to address this. This method uses a multivariate empirical model with iterative weights to determine the window adaptively. It is combined with Stacking-InSAR, a fast deformation inversion technology, for wide-area landslide detection and monitoring. Using a total of 140 Sentinel-1 SAR images, the reservoir area of the Wudongde hydropower station in Yunnan Province of China was selected as the study area. The experimental results showed that the standard deviation (STD) of the phase, after being corrected by MMEM, decreased by 55.25% compared to the initial values, reducing from 1.9078 to 0.8538. A total of 44 landslides were identified in the study area and 18 of which pose a threat to surrounding villages. The proposed method provides a rapid early warning workflow for the investigation of wide-area active landslides and supports geological disaster investigation departments.

Multi-Aspect Polarimetric Synthetic Aperture Radar (PolSAR) data captures the polarimetric properties of man-made buildings from various viewing angles, offering a more comprehensive understanding of scattering characteristics, valuable for target recognition and precise classification. To accurately characterize the multi-aspect and fully polarimetric SAR signatures of various man-made structures, separating anisotropic from isotropic scattering in the raw data is essential. However, the task of distinguishing between anisotropic and isotropic man-made targets poses challenges due to the potential similarities in polarimetric characteristics across different angles. Consequently, this paper introduces a change detection similarity matrix, constructed using the Likelihood Ratio Test (LRT) distance to perform azimuth sequence filtering. Subsequently, the Detecting-Removing-Incoherent-Adding (DRIA) framework for anisotropy scattering detection is employed to sequence the removal of anisotropic man-made buildings, with the remaining data undergoing incoherent integration for further analysis.

Experiments were conducted utilizing a dataset of airborne P-band full-polarimetric circular SAR data, acquired by the Aerospace Information Research Institute, Chinese Academy of Sciences, in the Mianyang area, Sichuan in 2008. Two different likelihood ratio detection methods, using constant false alarm rate (CFAR) detection based on the DRIA framework were utilized. Anisotropic man-made structures such as buildings and power lines within the experimental scene were correctly identified, along with their primary scattering directions. By comparing the detection results pre and post azimuth sequence filtering, we demonstrated that azimuth sequence filtering enhances the effectiveness and stability in the detection of anisotropic targets.

The phase error caused by orbital errors, atmospheric factors, and other factors can seriously affect the imaging effect of tomographic SAR. Therefore, it is necessary to correct the phase error in tomographic SAR data. This study proposes an improved autofocus phase calibration algorithm for tomographic SAR, which addresses the problems of traditional autofocus phase calibration methods such as dependence on initial value setting, poor robustness of evaluation indicators, and susceptibility to local optima. Firstly, external DEM data is used to separate and fit low order phase errors, enhancing the applicability of the autofocus algorithm to large amounts of phase error data; Secondly, two-dimensional image entropy was adopted as the optimization objective of the autofocus algorithm, which has stronger robustness compared to traditional information entropy and contrast indicators; In addition, using particle swarm optimization algorithm as the optimization algorithm solves the problem of traditional search methods relying on initial value setting and easily getting stuck in local optimal solutions. This study validated the improved method based on the airborne P-band tomographic SAR dataset from the Saihanba Forest Region in 2023. The experimental results show that compared to existing autofocusing phase calibration methods, the proposed method can better remove phase errors in tomographic SAR data, improve tomographic imaging quality, and enhance the accuracy of forest height extraction in tomographic SAR.

Long-wavelength Synthetic Aperture Radar (SAR) exhibits strong penetration capability through forest canopies, making it suitable for capturing observation information sensitive to forest structure. However, it is also vulnerable to interference of ground signal unrelated to the forest. Interferometric ground cancellation technique effectively attenuates sub-canopy ground scattering and enhances forest canopy scattering, resulting in improved sensitivity of extracted forest canopy intensity to Above-Ground Biomass (AGB). Nevertheless, this method relies on high-precision digital terrain model and cannot completely remove ground scattering contributions based on single-polarization data. To address this issue, our research proposes a polarimetric-enhanced interferometric ground cancellation algorithm. This method considers the removal of ground scattering from different scattering mechanisms using full-polarization data. By employing the SKP decomposition method, we extract volume scattering component and interferometric phase, enabling estimation of forest canopy backscatter. We validate the improved approach using airborne P-band PolInSAR data from the Saihanba forest area in 2023. Experimental results demonstrate that forest canopy backscatter intensity estimated by our proposed method has a better sensitivity to forest AGB compared to traditional ground cancellation method, with an increased R² value from 0.55 to 0.61.

Forests globally have experienced significant changes due to forest loss [7], underscoring the urgent necessity for real-time forest surveillance to mitigate further vegetation loss and enable timely interventions. Historically, forest monitoring relied heavily on optical imagery [3], which is impaired by its susceptibility to cloud coverage, particularly in tropical areas. In recent years, Synthetic Aperture Radar (SAR)-based systems have emerged, offering the advantage of all-weather operability [2], [5], [6]. However, SAR-based methods face challenges, such as changes in backscatter caused by factors like soil moisture variations. Moreover, accurately identifying small-scale disturbances remains problematic for SAR systems, partly due to spatial filtering techniques employed to reduce the effects of speckle. Furthermore, monitoring forest loss in regions with pronounced seasonality in backscatter signals, such as dry forests and savannas, presents limitations, leading to significant under-monitoring of these extensive carbon sinks.

This study presents an unsupervised SAR-based technique for identifying forest loss, utilizing Bayesian inference via an infinite state Markov chain. The methodology treats forest loss as a change-point detection problem within a Radiometrically Terrain Corrected (RTC) Sentinel-1 single polarization time series. Notably, this approach maintains the original resolution of measurements without employing spatial filtering. Each new observation contributes to the likelihood of deforestation, incorporating prior knowledge and a data model [1]. The method's iterative adaptation ensures robustness against fluctuations and trends, enabling forest loss monitoring not only in dense forests but also in regions affected by seasonal variations.

During the evaluation of the proposed method, tests were conducted by comparing various configurations, representing different levels of conservatism, against existing Near Real-Time (NRT) forest loss monitoring systems, including GLAD-L [3], RADD [6], and combined Global Forest Watch (GFW) alerts during the observation year 2020. Specifically, GFW incorporates alerts from GLAD-L, GLAD-S2, and RADD whenever available. The assessment primarily focused on small validation polygons (i.e., <1ha) in both the Brazilian Amazon and the Cerrado woodland savanna. Performance metrics such as detections, omissions, and false alarms were computed in comparison to the MapBiomas Alerta validation dataset [4].

Our research unveiled significant progress in detecting small-scale disturbances, accompanied by a remarkable reduction in false alarms across the examined biomes. In the Brazilian Amazon, our method achieved an F1-score of 97.3%, surpassing the 93.1% obtained by the current best-performing NRT system. Furthermore, a focused comparison underscored the existing systems' tendency to overestimate forest loss, likely due to spatial filtering impacting data resolution. In contrast, our method demonstrated more precise detections in absence of filtering and significantly lower false alarm rates in comparison to all considered systems, regardless of the configuration. In the Cerrado, our approach attained an F1-score of 97.4%, distinctly surpassing the 75.5% obtained through leading optical technology. In conclusion, our adaptive approach significantly improved forest loss detection with low false alarm rates, demonstrating efficacy in both the extensively monitored Amazon, and the Cerrado, where seasonal changes pose challenges to existing systems, leading to limitations or the absence of monitoring.

Our study revealed substantial advancements in detecting small-scale disturbances, coupled with a noteworthy decrease in false alarms across the investigated biomes. In the Brazilian Amazon, our method achieved an F1-score of 97.3%, outperforming the 93.1% achieved by the current best-performing NRT system. Additionally, a focused comparison highlighted the tendency of existing systems to overestimate forest loss, likely attributable to spatial filtering impacting data resolution. Conversely, our method demonstrated more accurate detections without filtering and notably lower false alarm rates compared to all considered systems, regardless of their configuration. In the Cerrado, our approach achieved an F1-score of 97.4%, significantly surpassing the 75.5% attained by leading optical technology. In summary, our adaptive approach significantly enhanced forest loss detection with minimal false alarm rates, exhibiting effectiveness in both the extensively monitored Amazon and the Cerrado, where seasonal variations challenge existing systems, leading to limitations or lack of monitoring.
References

[1] R. P. Adams and D. J. MacKay, Bayesian Online Changepoint Detection, arXiv 0710.3742, 2007.

[2] J. Doblas, . M. S. Reis, A. P. Belluzzo, C. B. Quadros, D. R. V. Moraes, C. A. Almeida, L. E. P. Maurano, A. F. A. Carvalho, S. J. S. Sant’Anna and Y. E. Shimabukuro, "DETER-R: An Operational Near-Real Time Tropical Forest Disturbance Warning System Based on Sentinel-1 Time Series Analysis," Remote Sensing, vol. 14, 2022.

[3] M. Hansen, A. Krylov, A. Tyukavina, P. Potapov, S. Turubanova, B. Zutta, S. Ifo, B. Margono, F. Stolle and R. Moore, "Humid Tropical Forest Disturbance Alerts Using Landsat Data," Environmental Research Letters, 2016.

[4] MapBiomas, Alert Project - Validation and Refinement System for Deforestation Alerts with High-Resolution Images, accessed in 2024.

[5] S. Mermoz, A. Bouvet, T. Koleck, M. Ballère and T. Le Toan, "Continuous Detection of Forest Loss in Vietnam, Laos, and Cambodia Using Sentinel-1 Data," Remote Sensing, vol. 13, 2021.

[6] J. Reiche, A. Mullissa, B. Slagter, Y. Gou, N. E. Tsendbazar, C. Odongo-Braun, A. Vollrath, M. Weisse, F. Stolle, A. Pickens, G. Donchyts, N. Clinton, N. Gorelick and M. Herold, "Forest disturbance alerts for the Congo Basin using Sentinel-1," Environmental Research Letters, 2021.

[7] C. Vancutsem, F. Achard, J. F. Pekel, G. Vieilledent, S. Carboni, D. Simonetti, J. Gallego, L. E. Aragão and R. Nasi, "Long-term (1990–2019) monitoring of forest cover changes in the humid tropics," Science Advances, vol. 7, 2021.

Obtaining the information of crop spatial distribution, type and growth timely and accurately plays a key role for achieving the optimal allocation of agricultural resources and ensuring national food security. Polarimetric synthetic aperture radar data (PolSAR) has unique advantages in dynamic monitoring crop growth, identifying crop type and obtaining crop spatial distribution since its sensitivity to the geometry and scattering direction of vegetation scatterers. In this study, we proposed a parcel-oriented crop types and their phenological phases identification using multi-temporal polarization SAR datasets. Wheat and rape parcels were taken as the research objects, and the experimental verification was carried out by using 5-phase Radarsat-2 PolSAR data and ground measured data. Firstly, parcel boundaries were extracted by a multi-temporal PolSAR image segmentation algorithm based on Mean-shift and Spectral Graph Partition (MS-SGP) theories. Then, taking parcels as the basic classification unit, a parameter optimization algorithm using genetic algorithm combined with support vector machine (GA+SVM) was used to identify crop types and phenological phases in multi-temporal PolSAR images. Next, the classification results using GA+SVR algorithm were compared with using traditional SVM parameter selection based on grid searching method (GA+ Grid SVM). The results show that the accuracy of the boundary extraction using MS-SGP algorithm reached 87.85%, which meets the needs of agricultural management. The crop-type classification accuracy using GA+SVM was 89%, which was 5.18% higher than that using GA+Grid SVM algorithm.The phenological phase recognition accuracy of wheat and rape is 87.88% and 88.75% respectively, which is 4.55% and 2.5% higher than that of using GA+Grid SVM algorithm. The results demonstrated the potential of proposed algorithms for parcel boundary extraction, crop type classification and phenological identification. The study provides a new solution and technical method for the automatic acquisition of agricultural information.

Accurate forest AGB estimation is the primary approach to quantify carbon stocks and sequestration rates. LiDAR signals can penetrate the forest canopy to obtain three-dimensional structural information from the top to the bottom of the forest, which benefit to understand the carbon sequestration capacity of forest ecosystems. Traditional forest AGB inversion methods based on LiDAR technology often requires setting up a large number of field sample plots, and it is difficult to obtain AGB results at the individual tree level. The LiDAR Biomass Index (LBI) has been proved to be able to complete accurate AGB estimation at individual tree level based on the crown size and tree height parameters obtained from terrestrial and airborne laser scanning data. More importantly, only a small number of sample trees were required to achieve model calibration of a certain tree species.

Unmanned Aerial Vehicle (UAV) LiDAR technology has the advantages of high accuracy, portability, and low cost, and has gradually become a commonly used data acquisition method in forest survey. This study aims to construct individual tree level AGB models for different tree species and tree species groups in China and European based on UAV LiDAR data, and verify their accuracy for further application. Firstly, 10 major coniferous and broad-leaved tree species around are selected, and the UAV LiDAR data of these tree species are acquired or collected. Secondly, the LiDAR data are segmented into individual trees and matched with field measurement data. Based on the matching results, a small number (≥30) of sample trees with different diameter classes for each tree species were selected. The LBI and tree height parameters of the sample trees are calculated based on the LiDAR data, and the AGB models for different tree species at individual tree level are constructed and verified in combination with the measured AGB. Then, the transferability between biomass models of similar tree species was evaluated and high-precision modelling methods for tree species groups are explored combine with the geometric characteristics of tree species.

The results indicate that LBI calculated from UAV LiDAR data can be used to construct the individual tree level AGB models for the 10 tree species (The R² of AGB models for different tree species range from 0.82 to 0.90). When selected individual trees to validate the accuracy of each model, the values of R² are between 0.72 and 0.88 and RMSE are between 45.32 kg and 282.90 kg for different tree species. Meanwhile, the biomass models have a certain degree of universality among tree species with similar geometric features, thereby obtaining acceptable modelling accuracy for tree species groups.The results of this study demonstrate the high accuracy and stability of LBI obtained from UAV LiDAR for individual tree level biomass calculation of different tree species, providing a new method for biomass calculation under “dual carbon” targets and laying a research foundation for its widespread application.

During the past two decades, advancements in artificial intelligence have ushered in new opportunities and challenges for integrating and innovating research in geospatial fields. Notably, the rise of deep learning models and development frameworks, and their application across specialized domains, have significantly accelerated the growth of Geospatial Artificial Intelligence (GeoAI). Nonetheless, GeoAI's advancement faces two major challenges: limited model interpretability and insufficient verification of interpretative outcomes. First, traditional "black box" models, when employed to tackle geospatial inference and prediction in Earth system science and social sciences, encounter issues related to opaque decision-making and lack of interpretability. Second, the burgeoning field of interpretable AI technologies addresses AI's interpretability issues but introduces new challenges for verifying and evaluating interpretive results, thus impacting model reliability. Enhancing the interpretability of traditional black box models and focusing on the evaluation of interpretive outcomes can thus bolster causal analysis of geographic events and phenomena, enriching our understanding, simulation, and prediction of geographic issues.

Identifying and interpreting the key drivers of grassland degradation/restoration remains a pivotal research challenge within the context of grassland degradation and recovery. This study leverages contemporary mathematical statistical methods and artificial intelligence technologies to pinpoint and elucidate the drivers of grassland degradation, a critical step for the recovery and sustainable development of grasslands in arid and semi-arid regions. It aims to thoroughly and systematically grasp the process and determinants of land degradation on the Mongolian Plateau. Given Xilingol's extensive grassland coverage and its exposure to nearly all ecological policies enacted in China, this area is chosen as a case study to identify and interpret the primary drivers behind grassland degradation and to gain a deep understanding of the intricate relationships between grassland degradation and its key drivers.

At the county scale, the primary drivers of grassland degradation were identified as follows: in the initial study period (1980-2000), land degradation emerged as the predominant land use change in Xilingol League, constituting 11.4% of the total area. During this time, human activities were the chief culprits of degradation in eight counties, whereas water resource conditions drove grassland degradation in six counties. In the subsequent phase (2000-2015), although the extent of land restoration expanded (representing 12.0% of the total area), degradation persisted, resulting in an additional 9.5% of land becoming degraded. In this phase, urbanization emerged as the primary driver of land degradation in seven counties, with the effects of human disturbance and water resources on grassland degradation diminishing post-2000.

After ascertaining the main drivers of degradation, integrating artificial intelligence methods with interpretable AI models facilitates a visual and quantitative exploration of the complex interactions between driving factors and grassland degradation. Pixel-scale findings reveal that proximity to grasslands of varying coverage levels and livestock density are key influencers of grassland degradation dynamics, with areas of high-quality grassland facing an elevated risk of degradation.

This research applies partial order theory and Hasse diagram techniques, considering the holistic and synergistic nature of driving factors, to discern the dominant drivers of land use change across different study periods within the selected area. It establishes a framework for identifying key driving factors based on a multi-index comparison model. Furthermore, through the use of interpretable artificial intelligence technology and the development of multi-level model interpreters for deciphering black box models, the study elucidates the logical decision-making process from "input (driving factors) to output (land use change)," delineates the nonlinear response patterns between land use change and driving factors, and achieves a comprehensive understanding of the driving mechanisms behind land use change in the study area at both local and global scales.

Over the years there has been a gradual deterioration of the planet's environmental conditions. Desertification and land degradation are among the most significant worldwide phenomena. They were long neglected until the United Nations Conference on Desertification (UNCOD) in 1977. After that, the comprehension of desertification and land degradation increased, with international interest in arid, semi-arid, and sub-humid areas of the world, whose importance for environment, food production and social development has been recognised. Desertification became a major topic for the “Life on Land” SDG15 of the 2030 Agenda for Sustainable Development and the Target 15.3 Land Degradation Neutrality. Grassland is the main component in arid, semi-arid, and dry sub-humid areas and is one of the most important and unique ecosystems in the world. It constitutes the largest terrestrial ecosystem and accounts for approximately 40% of the land (excluding Greenland and Antarctica). The grasslands provide countless services to the environment, animals, and humans from local to global scale. Some of these are climate regulation, carbon sequestration, soil and water conservation, sand and nitrogen fixation, nourishment for livestock, habitat for animals and recreation opportunities. The relationship between grassland and humans dates to ancient times and it is mainly based on the utilisation of the land and its resources for fuel, crops, and grazing. Over time, increasing use of the resources resulted in excessive exploitation with the gradual diminution of goods and services that these ecosystems can naturally provide. Human activities, such as overgrazing, are the main drivers of changes in vegetation and soil features. These lead to a loss of the land's ecological functions and are considered grassland degradation. Grassland degradation monitoring is imperative to expand the comprehension of desertification and to support policymaking, action plans and mitigating measures for the sustainable development of the land. Remote sensing is a technology that proves to be remarkably suitable for grassland degradation monitoring, bringing advantages in terms of efficiency, cost, and time. However, a review of the studies on grassland degradation monitoring with remote sensing exposes limitations and gaps in the literature. The studies use satellite images with a spatial resolution that does not allow the identification of land features in less than a few meters. This leads to contamination of the spectral information contained in a pixel, which consequently makes the analyses inaccurate. Also, it limits the observation of those degradation processes that occur at a small scale but that are still important components of the whole degradation mechanism. Grassland degradation is a complex phenomenon determined by the combination of vegetation and soil degradation and can be studied if all the attributes of desertification are specified. These are vegetation area, density, structure, and composition, soil water and wind erosion, soil compaction, soil waterlogging-salinisation-alkalinisation and rainfall variation. However, some studies assess grassland degradation using indicators that consider only vegetation degradation. Despite attempts of some more recent studies to take into account also soil degradation using soil indicators, mainly ground-based, grassland degradation is still underspecified. In recent years, new satellites made imagery at very high spatial resolution (VHR) available, which started to bring advances in land monitoring research. Yet, they have still to be explored for grassland degradation monitoring. Their resolution (lower than one meter) reduces the spectral contamination in a pixel and allows the identification of small vegetation and soil features. This can enhance the detail of the analyses and expand the consideration of the attributes of desertification. As a result, the detection and assessment of grassland degradation can be more consistent and comprehensive. This study explores the potential for using VHR satellite images for grassland degradation monitoring.

Grasslands are vital ecosystems, standing out for their complexity and significant role in supporting biodiversity, aiding in carbon sequestration, and offering livelihoods for various communities. The globe is covered by about 41% of grasslands that are home to more than 38% of people, mostly from low and medium-income countries. Overgrazing is considered the most significant cause of arid and semi-arid grassland degradation and desertification.

Existing research has made progress in using spectral indices to analyse critical features such as above-ground biomass, which is an important health indicator for grassland ecosystems. However, these studies do not adequately account for the multidimensional character of environmental degradation. A thorough assessment of grassland degradation, for example, would benefit from considering a broader range of sources and other physical indicators, such as soil degradation and microclimatic conditions, which interact with each other. This study proposes image fusion to achieve a comprehensive evaluation of grassland degradation, marking a pivotal advancement in environmental monitoring.

Image fusion involves the combination of various data types, such as optical and SAR imagery, enhancing our understanding of ecosystem health. For example, including SAR data, could potentially solve some of the constraints by offering new insights into the physical structure of the vegetation and soil moisture conditions and cloud-related issues. At the same time, optical sources provide information on the reflective and emissive characteristics of the surface. Moreover, Light Detection and Ranging (LIDAR) data sensors can capture the three-dimensional arrangement of vegetation canopies and the terrain beneath them, resulting in rich topographic maps and precise assessments of vegetation height, coverage, and canopy structure.

Recent research has taken a multi-level fusion approach to leverage the combined strengths of pixel-level and feature-level methods. This allowed for obtaining meaningful insights from remote sensing sources representing a comprehensive method for mapping vegetation degradation and soil degradation. The field remains underdeveloped with few dedicated studies, highlighting significant research opportunities. This also highlights that some significant ecosystem attributes might be overlooked and suggests a more comprehensive approach for matching sensor and phenomena characteristics.

To address these limitations, this study is devising a novel integration of area-focused and object-focused fusion to meet the challenge of assessing land degradation on both smaller and larger scales. Area-focused fusion leveraging sensors with wide coverage to produce a comprehensive view of grassland degradation. This captures expansive phenomena patterns without compromising the spectral integrity of the data. Object-focused fusion presents a detailed analysis of identified patterns that employs feature extraction and object-based image analysis techniques to segment images into discrete, identifiable objects. The integration of area-focused and object-focused fusion is a significant development in environmental monitoring. By connecting sensor properties with environmental hazards, such as the sensor/ bands' sensitivity to soil moisture, biomass, and vegetation structure, a thorough understanding of grassland degradation can be achieved.