Conference Agenda

Microwave and optical remote sensing data, including Sentinel-1, Sentinel-2, GF-3, Landsat-8, and ICESat-2, were utilized to analyze the salt field types, salt crust types, lake water depth, and sediment thickness in the Qarhan Salt Lake region. The classification and inversion capabilities of different data sources were comprehensively assessed. The application potential of polarimetric SAR data in salt lake monitoring was systematically evaluated, demonstrating its advantages in type identification, change detection, and inversion accuracy enhancement.

For multi-temporal polarimetric SAR (PolSAR) feature extraction, a time correlation entropy feature was proposed based on satellite-borne Sentinel-1 data. A multi-temporal correlation matrix of dominant scattering mechanisms was constructed, and the degree of target variation over the entire observation period was quantified using Shannon entropy. This approach overcomes the limitation of traditional multi-temporal features that only capture changes between two time points, enabling a more comprehensive representation of cumulative changes in surface cover types. Furthermore, by incorporating the principle of maximum entropy, a classification method based on interval estimation of distribution parameters was developed, which enhances spatial consistency and noise robustness. The proposed method was applied to feature extraction and classification in the Qarhan Salt Lake region, achieving superior performance compared to conventional PolSAR features and classification approaches.

To optimize polarization features, we addressed the challenge of high similarity in the polarimetric scattering characteristics of different salt crust types in the Qarhan Salt Lake. These features are predominantly governed by surface scattering and mixed with a small amount of volume scattering, resulting in considerable redundancy among features. To address this, a feature selection method based on polarimetric feature similarity metrics was developed using high-resolution GF-3 polarimetric SAR data. A statistical and textural similarity-based metric, referred to as the Statistical Similarity-based Feature Similarity Measure (SSFSM), was proposed to quantify inter-feature similarity. Based on this metric, seven representative polarimetric features were selected, effectively reducing feature dimensionality while maintaining high classification accuracy (>95%), which is comparable to the performance using the full feature set. This approach not only enhances the efficiency of salt crust classification but also provides a reliable technical foundation for subsequent high-precision identification and automated classification of salt lake surface features.

For fine classification of salt crusts, we proposed a feature optimization and reconstruction method based on neighborhood autoencoder (AE). This method addresses two major challenges: speckle noise interference in high-resolution polarimetric SAR images of homogeneous salt lake surfaces, and the high similarity and redundancy among the polarized scattering features of different salt crust types. The method reconstructs and compresses the original polarimetric features from fully polarized and dual-polarized SAR data, improving feature representation and alleviating redundancy. We performed classification validation experiments based on field survey results. The results show that the method is effective in extracting salt crust distributions and significantly enhances classification accuracy and efficiency.

Optical remote sensing demonstrates outstanding capabilities in water body monitoring. Therefore, research on the water depth inversion of salt lakes and the fine classification of salt fields based on optical remote sensing has been conducted, expanding the means for fine monitoring of salt lake resources. For water depth inversion inversion, a remote sensing method based on the multilayer perceptron (MLP) model was used. Time-series monitoring of salt lake depth changes was achieved through the combined use of multispectral and LiDAR data. Seven spectral bands from Landsat-8 L2 data were used as inputs, with ICESat-2 ATL13 data serving as labels for model training and fitting. This method provides an effective technical solution for salt lake water depth inversion. By integrating multispectral data with LiDAR, the inversion accuracy was improved, achieving a high accuracy (>85%). This approach provides strong support for predicting salt lake water depth and optimizing salt field processes.

In terms of salt field refined classification, Sentinel-2 multispectral data were used to classify and study the sediment thickness in carnallite pool areas. Four types of sediment situation were labeled based on actual collection data. To address the linear operation characteristics in carnallite pools, a line target recognition method incorporating the attention mechanism was proposed. Based on this, a semi-empirical and semi-structural line target classification model was developed. This model integrates linear geometric features and structural information, improving the accuracy of sediment thickness recognition in carnallite pools. Field investigations and comparisons demonstrated the feasibility of this method in distinguishing sedimentation levels, providing effective support for the timing of carnallite collection and resource distribution assessment.

At present, the inversion of salt crust roughness and water content based on PolSAR, estimation of regional salt content in salt fields, water depth changes using optical remote sensing, and monitoring of dynamic changes in sediment thickness are all included in the research plan. These objectives aim to further enhance the capability for high-precision inversion and quantitative monitoring of salt lake resource parameters.

The project aims to produce high-precision, high-resolution forest type classification maps, monitor forest growth dynamics using multi-temporal LiDAR data, invert forest vertical structure parameters from LiDAR data, and develop multi-scale forest quality evaluation methods. Currently, research has been conducted on forest parameter estimation based on TECIS data, optimization of forest type classification algorithms, and extraction of understory sapling parameters, providing critical data support for further analysis of forest ecosystems.

A forest parameter extraction method was developed based on TECIS data. By extracting waveform characteristics from TECIS LiDAR spot data and integrating airborne forest carbon stock data, a machine learning model was constructed to estimate forest carbon stocks at the spot scale. Subsequently, reflectance, texture, and BRDF features derived from multi-angle imagery were combined with spot-scale carbon stock estimates to build a deep learning model for regional forest carbon stock mapping. The mapping results were validated using both airborne and plot-level carbon stock data. Multi-scale accuracy assessments of the TECIS active-passive collaborative carbon stock maps were conducted using ALS-derived carbon data and field measurements, demonstrating high precision across different scales (plot scale: R²=0.45, RMSE=13.24 t/ha; 25m scale: R²=0.58, RMSE=11.74 t/ha; 100m scale: R²=0.70, RMSE=10.12 t/ha; 1km scale: R²=0.85, RMSE=5.18 t/ha).

A forest type mapping methodology integrating Sentinel-1/Sentinel-2 data with Digital Terrain Models (DTM) was developed. The approach utilized Sentinel-1's VV and VH backscattering coefficients, Sentinel-2's original spectral bands, three vegetation indices (NDVI, MSAVI2, and GEMI), along with three DTM-derived topographic variables (elevation in meters above sea level, aspect, and slope) to construct a random forest classification model for forest type discrimination. The resulting forest type map was validated against Switzerland's National Forest Inventory (NFI) plot data. Experimental results demonstrated that the synergistic use of Sentinel-1/Sentinel-2 and DTM effectively mitigates challenges posed by complex terrain and atmospheric conditions in mountainous regions. This model leverages freely available Sentinel-1/Sentinel-2 data and open-source software, featuring reproducible procedures with updatable training datasets. The implemented workflow proves robust, cost-efficient, and highly automated, significantly supporting forestry sector activities requiring precise spatial characterization of forest resources.

A LiDAR-based method for understory sapling parameter extraction was developed. The approach first optimized a spectral clustering-based individual tree segmentation algorithm to accurately delineate larch canopy trees and identify their trunk positions in the study area. After effectively removing overstory tree information, the method employed a mean shift algorithm with adaptive kernel bandwidth, incorporating local stem density, to segment understory spruce saplings. Subsequently, individual sapling height and crown phenotypic parameters were extracted from the segmented point clouds. The detection rate for understory saplings ranged from 94.41% to 152.78%, with matching accuracy improving from 62.59% to 95.65% as canopy closure decreased. Airborne Laser Scanning (ALS)-derived sapling heights showed strong correlation with field measurements (R²=0.71, N=3241, RMSE=0.26 m, p<0.01) and high consistency with Terrestrial Laser Scanning (TLS) data (R²=0.78, N=443, RMSE=0.23 m, p<0.01). ALS-extracted crown widths also demonstrated acceptable accuracy compared to TLS measurements (R²=0.64, N=443, RMSE=0.24 m). The results indicated systematic underestimation of sapling dimensions by ALS due to overstory occlusion effects. This research establishes a foundation for enhancing forest carbon sequestration capacity through regulating gap dynamics and optimizing light resource utilization in managed forests.

翻译

搜索

复制

The China-ESA Forest Observation (CEFO-2) project focuses on temperate and subtropical forests in China, and forest of temperate forests in United Kingdom. Recognizing the critical role of forests in carbon sequestration and the current limitations in quantifying carbon stocks and predicting trends, the project aims to conduct national-scale dynamic forest resource surveys by integrating ground-based and satellite remote sensing technologies. Over the 1^st year, CEFO-2 has achieved following results: (1) integration of multi-platform LiDAR and high resolution optical imageries for forest parameter extraction, and (2) forest dynamics classification and prediction.

In terms of multi-source data integration for forest parameter extraction, CEFO-2 utilized multi-platform LiDAR (terrestrial, airborne, spaceborne) and optical imagery to estimate key forest parameters—height, carbon stocks, volume, and DBH (diameter at breast height). Motivated by the limitations of legacy inventory systems such as the Sub-Compartment Database, the initiative addressed emerging forestry challenges including climate resilience, species diversity, and precision silviculture. A canopy detection algorithm was applied to delineate individual trees and generate detailed Treelists including height, crown dimensions, and spatial coordinates. Species-specific allometric models, developed using a combination of the global Tallo database and permanent sample plots in the UK, were used to estimate DBH from LiDAR-derived tree metrics. Results showed an exciting accuracy for coniferous species such as Douglas fir and Scots pine yielded high model accuracy (R² > 0.85), while broadleaved species showed greater structural variability, necessitating more nuanced modelling approaches. Volume estimates were calculated using species-calibrated equations, enabling accurate prediction of merchantable timber volume across the district. Complementary satellite-based species classification using PlanetScope imagery and Random Forest algorithms achieved over 92% accuracy by leveraging seasonal spectral variation. Integration of ALS, TLS, and satellite data facilitated the assignment of species attributes to individual trees, enabling the creation of a spatially coherent forest inventory. For spaceborne LiDAR, CEFO-2 acquired and processed spaceborne waveform LiDAR datasets of the Terrestrial Ecosystem Carbon Inventory Satellite (TECIS) from three representative temperate forest sites in northern China (Genhe site, Mengjiagang site, and Saihanba site) and across the UK. Using TECIS data, we extracted critical forest structural parameters such as vegetation height and carbon stocks.

In terms of forest dynamics classification and prediction, CEFO-2 firstly addressed the shortcomings of existing algorithms in disturbance type identification by employing a novel monitoring method based on Temporal Extended Convolutional Neural Network (TempCNN). This model employs 1D CNN to thoroughly analyze temporal dynamic features and spectral information from time-series remote sensing data. The classification result of Mengjiagang site indicated that the TempCNN can precisely classify the anthropogenic disturbance types including thinning, clear-cutting, canopy opening, selective canopy thinning, growth-enhancing silviculture, ecological thinning, forest closure and natural regeneration with the overall classification accuracy of 75%. The TempCNN is expected to provide a reliable tool for forest degradation attribution analysis. Secondly, this project focused on the improvement of forest management in the context of natural regrowth and managed plantation contexts. Work explored the integration of Digital Twin (DT) frameworks with artificial intelligence for improved forest management, particularly in natural regrowth and managed plantation contexts. A modular DT approach using remote sensing data, neural networks, and reinforcement learning was used to optimise forest operations such as thinning. A key contribution is the use of a custom reinforcement learning agent to simulate decision-making in silvicultural scenarios, demonstrating improved performance over traditional benchmarks. Complementing this, Physics-Informed Neural Networks (PINNs), incorporated biological growth models, specifically the Chapman-Richards equation, into the learning process to predict forest growth using sparse and noisy LiDAR data. This biologically grounded architecture significantly outperformed conventional data-driven models in robustness and generalisation, showing strong potential for use in DT systems. Together, these projects illustrated the viability of combining domain-specific ecological knowledge with machine learning to enhance forest simulation and prediction. Both highlight that embedding biological or procedural understanding into AI systems could improve interpretability and reliability. This interdisciplinary approach supported the UK's movement toward precision forestry practices, offering scalable digital tools for planning, carbon accounting, and climate resilience.

This year’s progress demonstrated that multi-source remote sensing integration enables spatially explicit forest inventory as Tree Lists and dynamic management strategies. Future efforts will strengthen the China-Europe scientific collaboration network, developed methods for monitor changes, improving species classification and the upscaling to estimates at a national level by engaging both Chinese and British datasets. The project will to facilitate policy applications of research outcomes, while cultivating interdisciplinary young researchers through project participation.