Conference Agenda

Forests play a critical role in the Earth's ecosystem and strongly impact the environment. Under the threat of global climate change, remote sensing techniques can provide information for a better understanding of the forest ecosystems, early detection of forest diseases, and both rapid and continuous monitoring of forest disasters. This project concerns the topic of ecosystems and spans the subtopics estimation of forest quality parameters and forest and grassland disaster monitoring. The aim is to study and explore the application of multi-source remote sensing technology in forest parameter extraction and forest disaster monitoring using data fusion of satellite images, drone-based laser scanning and drone-based hyperspectral images. The research contents include tree species classification, forest parameters estimation, and forest disturbance detection.

1. Work performed

(1) Satellite image data

We applied for satellite images through ESA and MOST of China, including RADARSAT-2 (2020 and 2021), WorldView-3 (June 2021), Sentinel-1/2 (from 2018 to 2023), and Gaofen-1/2/6 (from 2020 to 2023). These data cover several study areas including Gaofeng, Weihai, Fushun, Lu'an, Wangyedian, Genhe and Pu'er in China and Remningstorp in Sweden.

(2) Field investigation data

For different research contents, field investigations were carried out in Gaofeng, Fushun, Lu'an, Genhe, Pu'er and Remningstorp. The details are as follows:

l 2019：The forest information of the sample plots in Remningstorp, Sweden was updated.

l 2021: The forest information of the sample plots in Gaofeng and Genhe in China was updated; Spectral information from healthy and pine nematode-infested forests at different stages of the Fushun and Lu'an study areas in China was collected; The forest information of the sample plots in Remningstorp, Sweden was updated.

l 2023: The forest information of the sample plots in Gaofeng and Genhe in China was updated; Forest tree species types, forest changes and disturbance information of Pu'er study area in China were collected; The occurrence status and geographical distribution of Simao pine bollworm pests and diseases were recorded; Controlled experiments were conducted for bark beetle infestation. A weekly inventory of tree symptoms of bark beetle infestation was conducted. Four sapflow stations were built to monitor the hourly physiological status of 32 trees throughout the entire vegetation season. Multispectral and hyperspectral images were collected weekly.

(3) Technical progress

l Tree species classification. We proposed five pixel-based deep learning tree species classification models using drone-based hyperspectral data: an improved prototype network (IPrNet), a CBAM-P-Net model of the prototype network combined with an attention mechanism, a Proto-MaxUp+CBAM-P-Net model of the CBAM-P-Net combined with a data enhancement strategy, SCL-P-Net introducing contrast supervised learning and FAST 3D-CNN P-Net Model Combining Band Selection and FAST 3D-CNN Network for P-Net Backbone Networks. We evaluated and screened low-cost and efficient UAV optical image acquisition solutions for individual tree species identification，and developed an instance segmentation algorithm, ACE R-CNN, for individual-tree species identification using UAV LiDAR and RGB images. The performance of these models was demonstrated in the Gaofeng study area. A tree species classification method based on multi-temporal Sentinel-2 data was developed and the performance was verified at Remningstorp.

l Forest parameters extraction. We proposed a method for extracting crown parameters considering inter-tree competition using terrestrial close-range observation data with missing canopy information. We proposed a mean-shift individual-tree crown segmentation algorithm based on canopy attributes using UAV oblique photography data, and developed an individual-tree biomass estimation model fusing multidimensional features. A three-level stratified feature screening method fusing airborne hyperspectral and LiDAR data was innovated to construct regional AGB estimation models for different tree species. A high spatial resolution tree height extraction method combining ZY-3 stereo images and DEM was proposed, and a forest AGB estimation model using Sentinel-2 data and tree height data was developed to obtain accurate forest AGB maps in the Wangyedian study area. We proposed a quantitative method for thinning and clear-cutting phase height for detecting silvicultural treatment using the phase-height data from time-series TanDEM-X. In addtion, we investigated the use of interferometry (InSAR) of TanDEM-X images for estimation of forest changes (height, biomass and biomass change), and mapped smaller forest height changes (increase) in a boreal forest in Sweden.

l Forest disturbance detection. For Bursaphelenchus xylophilus, we analyzed the spectral characteristics of two tree species (Pinus tabulaeformis and Pinus koraiensis) in the study areas of Weihai and Fushun during different infection stages. Exploring the most informative wavelengths in unmanned aerial vehicle (UAV) hyperspectral images for early identification of pine wood nematode disease; Proposing an improved Mask R-CNN instance segmentation method and an integrated method combining a prototype network classification model and a single-wood segmentation algorithm to efficiently segment pine wood nematode infected wood at different dry infection stages; A conifer information extraction index (NDFI) based on time-series Landsat images was constructed. For European spruce bark beetles (Ips typographus [L.]) infestation, methods of early detecting infestations were proposed using drone-based multispectral images. We investigated how early the infestation can be detected after an attack. We also compared the machine-learning- and vegetation-index-based methods for the early detection of bark beetle infestations, and found the machine-learning-based methods had overfitting issues. We are now working on the uncertainties of the detectability in the early stage, using multispectral data with four-year time series on 6 different areas. We expect to quantify the variation of vitality decline and explain it using environmental factors and models. We also investigated the sensitive bands of hyperspectral drone images on early infestation and proposed modified indices for early detection. For forest disturbance, a CCDC disturbance detection algorithm incorporating spectral indices and seasonal features was proposed to robustly map forest disturbances over the past 30 years in the Genhe study area.

(4) Collaborative Research

l Co-supervising 1 PhD student.

l One academic paper in Ecological Indicators and one in Computers and Electronics in Agriculture. One conference papers were published in IGARSS 2022, and one joint conference paper was published in IGARSS 2023. One joint conference paper was under view by IGARSS 2024.

l In the process of building a Joint Laboratory for International Cooperation on Intelligent Remote Sensing Monitoring of Forests.

The airborne multi-dimensional SAR flight experiment in the forest area of Genhe district was organized in 2021, and the PolSAR dataset of P, L, S, C and X bands, P-band TomoSAR dataset and dual antenna InSAR dataset of C band were obtained. The P-band SAR flight experiment in the forest area of Zhaoqing and SaiHanBa were organized in 2022 and 2023, and the P-band TomoSAR dataset were obtained. Based on above dataset, we analyzed and evaluated the performance of PolSAR data of 5 bands and different band combinations in estimating forest volume, and evaluated the application effect of TomoSAR technology in different forest areas of China.

In the case of PolSAR data radiometric calibration method development, using space-borne GF-3 data and airborne UAVSAR data, we proposed the cross-co-polarization radio coefficient, which can be used to obtain the truth value of polarization scattering of any distributed targets. The calibration method can effectively reduce the constraints on targets of existing methods. In case of terrain radiometric correction (RTC), one RTC method for PolSAR based on RPC model has been proposed, which reduces the technical threshold for geometric and radiometric correction of PolSAR. In addition, a RTC method suitable for supervised classification of PolSAR was proposed, which can improve the accuracy of forest type classification by about 20%.

A series of land cover classification method studies were carried out using spaceborne Radarsat-2 and airborne UAVSAR time-series polarimetric SAR data. We constructed the time-polarization features, reflected the degree of feature variance, and constructed the foundation for effective feature selection; proposed the polarimetric and time dimension feature selection algorithms IESSM and SSV, designed a classifier based on Transformer, and reduced the feature redundancy and enhanced the adaptability of the classifier. We constructed a data representation power model, 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; extracted the time-variant scattering features based on the proposed power model, enhanced the expression ability of scattering variation, and improved the classification accuracy.

In terms of InSAR, the multi-layer model suitable for short wavelength InSAR is innovated, and the retrieval accuracy of forest height is improved by realizing that the observation calculation and theoretical model of InSAR coherence obey the same assumption. Moreover, the algorithm for jointly measuring forest height and forest AGB using P/X dual-frequency InSAR has been proposed, which effectively improves the extraction accuracy of DTM/DSM/CHM and AGB in forest areas.

Most current TomoSAR methods use a local means value of the sample covariance matrix, which may get the poorly refined spectrum, and lose some detailed information. In addition, the spectrum will inevitably produce sidelobe effects. To address the above issues, a non-local means method is applied to identify neighboring pixels with high similarity to the target pixel, thereby comprehensively reflecting its feature information. Moreover, G-Pisarenko method is introduced in TomoSAR to reduce the spurious interference signal. These two methods have been respectively verified that the feasibility and effectiveness with BioSAR 2008 L-band data and AfriSAR 2016 P-band datasets, respectively. In addition, we propose a TomoSAR algorithm based on atomic norm minimization(ANM) to solve the scatterer location error caused by elevation discretization in traditional TomoSAR methods. The performance of the algorithm has been verified by TerraSAR-X data. TomoSAR baseline correction and phase compensation methods for multi baseline interferometric SAR assisted by DEM have been developed, improving the imaging quality of TomoSAR over forested areas. Additionally, we have developed a multi feature collaborative forest biomass estimation method based on TomoSAR profile fitting, which has achieved high accuracy (>90%) in tropical rainforest regions.

PolTomoSAR techniques have been developed to characterize tropical forests at P band using adaptive parametric signal processing approaches, and over temperate forests at L band using a minimal number of images. The benefits of a synergistic use of the different modes of the upcoming BIOMASS missions have been evaluated by computing the ultimate performance limits of this mission for different forest characteristics, and according to various temporal scenarios. The gain provided by external sources of information, such as GEDI, has been evaluated by coupling this estimation with Bayesian principles. The case of estimation techniques using models of forest vertical profiles that differ from actual ones has been investigated. Some of the developed techniques will contribute to the BIOMASS processing group of methods Multi-Mission Algorithm and Analysis Platform (MAAP). Some work also been done regarding the definition of BOMASS level 3 product processing chain, i.e. Forest Height, Above Ground Biomass and Forest Disturbance.

Times series of Sentinel 1 measurements were used for mapping deforestation at large scale (see https://www.tropisco.org/ ) and new techniques, based on Bayesian processing are being developed. A novel near-real time online forest loss detection technique allowed to reach very good results, compared with operational existing methods, such as Global Forest Watch, over both Amazonian forests and Savannah such as Cerrado, in Brazil.