Science of Remote Sensing最新文献

Monitoring forest growth accurately is important for assessing and controlling forest carbon stocks that impact, for example, the atmospheric CO₂ concentration and, consequently, the climate change. In prior studies, forest growth monitoring with laser scanning methods has resulted in relatively high errors. However, the contribution of reference measurement error to uncertainty in growth resolution has rarely been analysed, and the reference measurements are usually considered mostly flawless. In this study, a seven-year-long growth of individual trees was estimated using both airborne and terrestrial laser scanning (ALS, TLS) that have emerged as potential candidates for digital forest reference measurements. The growth values were derived for diameter at breast height (DBH) and stem volume between the years 2014 and 2021 using an indirect approach. The values obtained with laser scanning were paired with manual field measurements and also with each other to study pairwise errors. The pairwise comparison showed that even though all the three measurement methods produced good Pearson correlation coefficients for one-time measurements (all above 0.88), the coefficients for growth measurements were significantly lower (0.19–0.44 for DBH and 0.47–0.66 for stem volume). The best correlation and root mean squared deviation (RMSD) for DBH growth (ρ = 0.44, RMSD = 0.98 cm) and stem volume growth (ρ = 0.66, RMSD = 0.052 m³) was observed between the manual field measurements and the ALS-based growth measurement method, in which the tree stem curve was obtained from the 2021 point cloud, and the stem curve was predicted backwards for the year 2014 according to height growth. The ALS method suffered less from outlying values than the TLS-based growth measurement method, in which the growth was computed based on the difference of stem curves derived separately for the years 2014 and 2021. The study showed that observing the stem curve is a potential method for short-period growth monitoring. Using the pairwise comparison results, we further derived estimates for the mean and standard deviation of measurement error of each individual measurement method. For the manual measurements, the standard deviation of error was found to be approximately 0.4 cm for DBH growth and 0.03 m³ for volume growth, which were the lowest of the three methods but not by a large margin. This highlights the need for more accurate reference data as the accuracy of laser scanning-based growth estimation methods continues to approach the accuracy of manual measurements.

In order to reveal the spatial and temporal distribution of COVID-19's economic impact on the Beijing-Tianjin-Hebei region, this study uses the NPP/VIIRS night light remote sensing data from January to September in 2020 to compare the development trend of COVID-19 and analyze its economic impact on the Beijing-Tianjin-Hebei region. At the same time, the regional economic resilience measurement algorithm is introduced by coupling the regional night light greyscale value to obtain the economic resilience data of various cities during the epidemic. The findings show that: 1. there are structural differences in the spatial distribution of COVID-19 outbreaks in the Beijing-Tianjin-Hebei region. Beijing-Tianjin-Hebei region present a "core-adjacent-external" structure and the spatial distribution pattern of Tianjin-Beijing-Shijiazhuang prominent in the inverted "L" shape. 2. There are differences in the economic resilience of the Beijing-Tianjin-Hebei region in the face of the epidemic, with high economic resilience in the core urban areas close to Beijing and Tianjin. Therefore, strengthening regional cooperation and establishing relatively stable economic ties with surrounding areas are the key to improving the overall economic resilience of Beijing-Tianjin-Hebei region.

This study developed a random forest approach for downscaling the coarse-resolution (36 km) soil moisture measured by The National Aeronautics and Space Administration (NASA) Soil Moisture Active Passive (SMAP) mission to 1 km spatial resolution, utilizing airborne remotely sensed data (radar backscatter and radiometer retrieved soil moisture), vegetation characteristics (normalized difference vegetation index), soil properties, topography, and ground soil moisture measurements from before the launch of SMAP for training a random forest model. The 36 km SMAP soil moisture product was then downscaled by the trained model to 1 km resolution using the information from SMAP. The downscaled soil moisture was evaluated using airborne retrieved soil moisture observations and ground soil moisture measurements. Considering the airborne retrieved soil moisture as a reference, the results demonstrated that the proposed random forest model could downscale the SMAP radiometer product to 1 km resolution with a correlation coefficient of 0.97, unbiased Root Mean Square Error of 0.048 m³ m⁻³ and bias of 0.016 m³ m⁻³. Accordingly, the downscaled soil moisture captured the spatial and temporal heterogeneity and demonstrated the potential of the proposed machine learning model for soil moisture downscaling.

Woody plants in both forested and non-forested areas are vital for carbon storage, climate change mitigation, biodiversity conservation, and provision of ecosystem services. Accurate mapping of woody cover (WC) is crucial for understanding global environmental dynamics, but despite advancements in Earth observation (EO), challenges persist in WC mapping, particularly in spatially heterogeneous mixed tree-grass systems, characterized by low density and low stature (LDLS, i.e., savannas and dryland ecosystems) woody plants.

This study aims to guide users in selecting appropriate WC products for their analytical needs, particularly in LDLS ecosystems, and encourage WC product developers to consider incorporating dryland woody vegetation into their product development, utilizing modern EO data and techniques. To achieve this, we assessed existing WC products for the biome diverse Sub-Saharan Africa (SSA), for epoch 2005–2010 (EP01) and 2015–2020 (EP02). Our analysis focused on LDLS, which are often overlooked in EO products. We provide error assessments for available WC products at continental and regional scales, in both epochs, providing data for optimal dataset selection. Our results show that WC products that exclude low stature woody vegetation (<5 m height) from training data tend to underestimate WC in drylands, particularly in areas where WC is <40%. However, in general models tend to underestimate cover in dense WC ecosystems. This could potentially be attributed to systematic bias in machine learning regression models, lack of sufficient training data, and increased prevalence of cultivation, and cloud contamination in more humid regions.

In this paper, we compare the positioning accuracy of commercial, mobile laser scanning systems operating under a forest canopy. The accuracy was evaluated on a 800-m-long positioning track, using tree locations from both a traditional field reference, collected with total station, and a high-density airborne laser scanning (ALS) system as a reference. Tree locations were used since mobile lasers are studied for automation of field reference for forest inventory and location of individual trees with high accuracy is required. We also developed a novel method for evaluating the ground level around the trees, as it not only affects the z-coordinate, but the horizontal position as well if the tree is tilted.

In addition to the accuracy that could only be evaluated for systems equipped with a GNSS receiver, we evaluate the consistency of laser scanning systems by registering the tree locations extracted from the mobile systems to both the field reference and ALS. We demonstrated that the high-density ALS has similar accuracy (RMSE of approximately 6 cm) and precision as the total station field reference, while being much faster to collect. Furthermore, the completeness of the high-density ALS was over 80 %, which is more than enough to register the other methods to it in a robust manner, providing a global position for laser scanners without an inherit way of georeferencing themselves, such as a GNSS receiver.

The positioning of all the mobile systems were based on the Simultaneous Localization and Mapping (SLAM) algorithm integrated with an inertial measurement unit (IMU), and they showed a similar precision; planar positioning error of less than 15 cm and vertical error of 10–30 cm. However, the accuracy of the only commercial system in this test whose positioning methods included a GNSS receiver, was order of several meters, indicating a demand for better methods for GNSS-based global positioning inside a dense forest canopy.

For Landsat land cover classification, the time series observations are typically irregular in the number of observations in a period (e.g., a year) and acquisition dates due to cloud cover variations over large areas and acquisition plan variations over long periods. Compositing or temporal percentile calculation are usually used to transform the irregular time series to regular temporal variables so that the machine and deep learning classifiers can be applied. Recognizing that the composite and percentile calculations have information loss, this study presents a method directly Classifying the Raw Irregular Time series (CRIT) (‘raw’ means irregular good-quality surface reflectance time series without any composite or temporal percentile derivation) by adapting Transformer. CRIT uses the acquisition day of year as classification input to align time series and also takes the Landsat satellite platform (Landsat 5, 7 and 8) as input to address the inter-sensor reflectance differences.

The CRIT was demonstrated by classifying Landsat analysis ready data (ARD) surface reflectance time series acquired across one year for three years (1985, 2006 and 2018) over the Conterminous United States (CONUS) with both spatial and temporal variations in Landsat availability. 20,047 training and 4949 evaluation 30-m pixel were used where each pixel was annotated as one of seven land cover classes for each year. The CRIT was compared with classifying 16-day composite time series and temporal percentiles and compared with a 1D convolution neural network (CNN) method. Results showed that the CRIT trained with three years of samples had 1.4–1.5% higher overall accuracies with less computation time than classifying 16-day composites and 2.3–2.4% higher than classifying temporal percentiles. The CRIT advantages over 16-day composites were pronounced for developed (0.05 F1-score) and cropland (0.02 F1-score) classes and for mixed or boundary pixels. This was reasonable as the 16-day composites had only on average 7.02, 16.49 and 15.78 good quality observations for the three years, respectively, in contrast to 7.89, 27.72, and 26.60 for the raw irregular time series. The CNN was not as good as CRIT in classifying the raw irregular time series as CNN simply filling temporal positions with no observations as zeros while the CRIT used a masking mechanism to rule out their contribution. The CRIT can also take the pixel coordinates and DEM variables as input which further increased the overall accuracies by 1.1–2.6% and achieved 84.33%, 87.54% and 87.01% overall accuracies for the 1985, 2006 and 2018 classifications, respectively. The CRIT land cover maps were shown consistent with the USGS Land Change Monitoring, Assessment, and Projection (LCMAP) maps. The developed codes, training data and maps were made publicly available.

Microbial communities are the primary drivers of carbon cycling in the McMurdo Dry Valleys of Antarctica. Dense microbial mats, consisting mainly of photosynthetic cyanobacteria, occupy aquatic areas associated with streams and lakes. Other microbial communities also occur at lower densities as patchy surface biological soil crusts (hereafter, biocrusts) across the terrestrial landscape. Multispectral satellite data have been used to model microbial mat abundance in high-density areas like stream and lake margins, but no previous studies have investigated the lower detection limits of biocrusts. Here, we describe remote sensing and field-based survey and sampling approaches to study the detectability and distribution of biocrusts in the McMurdo Dry Valleys. Using a combination of multi- and hyperspectral tools and spectral linear unmixing, we modeled the abundances of biocrust in eastern Taylor Valley. Our spectral approaches can detect low masses of biocrust material in laboratory microcosms down to biocrust concentrations of 1% by mass. These techniques also distinguish the spectra of biocrust from both surface rock and mineral signatures from orbit. We found that biocrusts are present throughout the soils of eastern Taylor Valley and are associated with diverse underlying soil communities. The densest biocrust communities identified in this study had total organic carbon 5x greater than the content of typical arid soils. The most productive biocrusts were located downslope of melting snowpacks in unique soil ecosystems that are distinct from the surrounding arid landscape. There are similarities between the snowpack and stream sediment communities (high diversity of soil invertebrates) as well as their ecosystem properties (e.g., persistence of liquid water, high transfer of available nutrients, lower salinity from flushing) compared to the typical arid terrestrial ecosystem of the dry valleys. Our approach extends the capability of orbital remote sensing of photosynthetic communities out of the aquatic margins and into the drier soils which comprise most of this landscape. This interdisciplinary work is critical for measuring and monitoring terrestrial carbon stocks and predicting future ecosystem dynamics in this currently water-limited but increasingly dynamic Antarctic landscape, which is particularly climate-sensitive and difficult to access.

Soil organic carbon (SOC) plays a major role in the global carbon cycle and is an important factor for soil health and fertility. Accurate mapping of SOC and other influencing parameters are crucial to guide the optimization of agricultural land management to maintain and restore soil health, to increase soil fertility, and thus to quantify its potential for sequestering CO₂. Remote sensing and machine learning techniques offer promising approaches for predicting SOC distribution. In this study, we used remote sensing data and machine learning algorithms to map SOC at regional to large scale, which we then combined with temporospatial and spectral signature-based soil sampling to integrate local ground measurements. A rigorous validation approach was performed where several independent unseen datasets with a high number of samples were used, which additionally involved densely sampled fields. We found that our approach could predict SOC with an average percentage error of less than 10 % with an R² of 0.91 using support sampling on croplands located on mineral soils, demonstrating the potential of remote sensing, machine learning, and specific ground measurements for mapping SOC. Our results suggest that this approach could make small carbon differences measurable and inform carbon sequestration efforts and improve our understanding of the impacts of land use and field management practices on soil carbon cycling.

Applying remote sensing models outside the temporal range of their training data, referred to as temporal model transfer, has become common practice for large area monitoring projects that extrapolate models for hindcasting or forecasting to time periods lacking reference data. However, the development of appropriate validation methods for temporal transfer has lagged behind its rapid adoption. Breaking temporal transfer's assumption of temporal consistency in both remote sensing and reference data and their relationship to each other could lead to biased pixel-level predictions and small area estimators, compromising the operational validity of large area monitoring products. Few studies using temporal transfer have evaluated its effects on model accuracy at the pixel/plot level, and the propensity for biased small area estimators and trends resulting from temporal transfer remains unexplored. We present a framework for evaluating temporal transferability by combining temporal cross-validation with a multiscale map assessment to aid in identifying where and when biased model predictions could scale to small area estimates and affect predicted trends.

This validation framework is demonstrated in a case study of annual percent tree canopy cover mapping in Michigan. We tested and compared temporal transferability of random forest models of canopy cover derived from 2010 to 2016 systematic dot-grid photo-interpretations at Forest Inventory and Analysis plots with Landsat spectral indices fit with the LandTrendr temporal segmentation algorithm serving as the primary predictor variables. The temporal cross-validation error (mean RMSE = 13.9% cover) was higher than the common validation approach of considering all time periods of testing data together (RMSE = 13.6% cover) and lower than models trained and tested within the same year (mean RMSE = 14.2% cover). However, the bias of model predictions and small area estimators for individual years was higher with temporal transfer models than when applying models within the same year as their training data. We also evaluated how training models using different temporal subsets and with and without LandTrendr fitting affected predictions of Michigan's 1984–2020 predicted annual mean cover. The mean cover from LandTrendr-based models followed expected and consistent trends and had less difference between models trained with different temporal subsets (max difference = 5.8% cover). While those from Landsat had high interannual variations and greater difference between temporal models (max difference = 11.2% cover). The results of this case study demonstrate that evaluation of temporal transferability is necessary for establishing the operational validity of large area monitoring products, even when using time series algorithms that improve temporal consistency.

Several studies utilized C-band Synthetic Aperture Radar (SAR) backscatter time series to detect cut events of grasslands. They identified several potential factors hindering the detection: Vegetation characteristics, precipitation, and the timing of salvage of the harvested grass. This study uses a comprehensive in situ database to assess the impact of those factors on the detection rate of cut events by performing a cut detection based on Sentinel-1 backscatter time series and relating the accuracy to the potentially limiting factors. The results can be summarized in the following key findings: (i) The detection rate decreases significantly with grass heights below 35 cm and a biomass of less than 2100 kg/ha. As the grass of the first growth is typically characterized by greater height and higher biomass, first cuts achieved a higher accuracy with 85% compared to re-growth cuts with 65%. (ii) False positive cut events were related to higher precipitation amounts, but adding precipitation data to the model led only to a slight increase of the accuracy of re-growth cuts, but a decrease of the overall accuracy. (iii) No relation was found between the timing of salvage and the backscatter behaviour. These insights contribute to a better utilization of C-band backscatter for vegetation analysis and agricultural applications, including cut detection. Further research with dense in situ measurements, including Vegetation Water Content (VWC) is required to fully understand the behaviour of C-band backscatter over managed grasslands.