Science of Remote Sensing最新文献

In recent decades, Earth Observation (EO) satellite missions have improved in spatial resolution and revisit times. These missions, traditionally government-funded, utilize state-of-the-art technology and rigorous instrument calibration, with each mission costing millions of dollars. Recently, nano-satellites known as CubeSats are presenting a cost-effective option for EO; their capacity of working as a constellation has brought an unprecedented opportunity for EO in terms of achievable spatial and temporal resolutions, albeit at the cost of decreased accuracy and cross-sensor consistency. As such, CubeSat datasets often require post-calibration approaches before using them for scientific applications. K-sharp is a relatively simple, data-agnostic machine learning approach that combines K-means and partial least squares regression to derive relationships between two sets of images for normalization. This study used Planet's four-band CubeSat imagery to sharpen day-coincident Landsat 8 normalized difference vegetation index, albedo, and the first short-wave infrared (SWIR) band from 30 m to 3 m spatial resolution (it should be noted that the four-band CubeSat product does not include the first SWIR band, and that the calculation of albedo is not directly possible from this product). K-sharp was tested over agricultural, savanna, rainforest, and tundra sites with and without atmospheric correction. Our model reproduced surface conditions with an average r² of 0.88 (rMAE = 11.39%) across all study sites and target variables when compared against the original Landsat 8 data. These results showcase the promising potential of K-sharp in generating precise, CubeSat-derived datasets with high radiometric quality, which can be incorporated into agricultural or ecological applications to enhance their decision-making process at fine spatial scales.

Accurate information on tree species diversity is critical for forest biodiversity, conservation and management, but mapping forest diversity over large and mixed forest areas using satellite remote sensing data remains a challenge because of scale- and ecosystem-dependent relationships between spectral heterogeneity and tree species diversity. In this study, three different diversity indices (Simpson (λ), Shannon (H’), and Pielou (J’)), were tested to characterize forest tree species diversity using individual monthly and multi-temporal Sentinel-1 and -2 images during 2021. The performance of three different machine learning models, Random Forest (RF), Extreme Gradient Boosting (XGB), and Deep Neural Network (DNN) were tested. A collection of 1,020 plot measurements (comprising 47 tree species and 28,122 trees), randomly collected in a mixed broadleaf-conifer forest area in northeast China, was used to train (n = 816) and validate (n = 204) the models. The models dependent on multi-temporal Sentinel-1/2 imagery were found to outperform the models based on individual monthly data, in predicting forest tree species diversity, with average accuracies of 78% for H’, 77% for λ and 77% for J’. The use of DNN performed marginally better than the XGB and RF models, with accuracies of 81% for H’, 80% for λ and 79% for J’, respectively. Finally, a boosted regression model, involving environmental variable predictors and the DNN-based estimated tree species diversity, showed that on average 63 ± 4% of the spatial variations of tree species diversity was explained by environmental variables, including annual temperature (29.30%), followed by soil fertility (27.03%), snow cover (13.63%) and a digital elevation model (12.33%). Our results highlight that an empirical approach based on machine learning and multi-temporal Sentinel-1/2 data can accurately predict forest tree species diversity and we further show the important roles of air temperature and soil fertility in governing the spatial variability of tree species diversity in a mixed broadleaf-conifer forest setting.

Forest structure is an essential part of biodiversity and ecological analysis and provides crucial insights to address challenges in these areas. Modern sensor technologies unlock new possibilities for more advanced vegetation monitoring. This study examines the potential of single high resolution X-band synthetic aperture radar (SAR) and optical images for pixel-wise mapping of four forest structure attributes (height, average height, fractional cover, and density) at a striking 0.5 m resolution. The study site is situated in Western Norway, hosting trees from flatlands to elevated mountainous areas and in-between. The proposed model architecture, called PSE-UNet, is a modified UNet incorporating key components from state-of-the-art deep learning from the field of forest structure monitoring. A comparative analysis involving state-of-the-art models shows promising results with MAE% between 21.5 and 24.7, depending on the variable.

Northern forest ecosystems make up an important part of the global carbon cycle. Hence, monitoring local-scale gross primary production (GPP) of northern forest is essential for understanding climatic change impacts on terrestrial carbon sequestration and for assessing and planning management practices. Here we evaluate and compare four methods for estimating GPP using Sentinel-2 data in order to improve current available GPP estimates: four empirical regression models based on either the 2-band Enhanced Vegetation Index (EVI2) or the plant phenology index (PPI), an asymptotic light response function (LRF) model, and a light-use efficiency (LUE) model using the MOD17 algorithm. These approaches were based on remote sensing vegetation indices, air temperature (T_air), vapor pressure deficit (VPD), and photosynthetically active radiation (PAR). The models were parametrized and evaluated using in-situ data from eleven forest sites in North Europe, covering two common forest types, evergreen needleleaf forest and deciduous broadleaf forest. Most of the models gave good agreement with eddy covariance-derived GPP. The VI-based regression models performed well in evergreen needleleaf forest (R² = 0.69–0.78, RMSE = 1.97–2.28 g C m⁻² d⁻¹, and NRMSE = 9–11.0%, eight sites), whereas the LRF and MOD17 performed slightly worse (R² = 0.65 and 0.57, RMSE = 2.49 and 2.72 g C m⁻² d⁻¹, NRMSE = 12 and 13.0%, respectively). In deciduous broadleaf forest all models, except the LRF, showed close agreements with the observed GPP (R² = 0.75–0.80, RMSE = 2.23–2.46 g C m⁻² d⁻¹, NRMSE = 11–12%, three sites). For the LRF model, R² = 0.57, RMSE = 3.21 g C m⁻² d⁻¹, NRMSE = 16%. The results highlighted the necessity of improved models in evergreen needleleaf forest where the LUE approach gave poorer results., The simplest regression model using only PPI performed well beside more complex models, suggesting PPI to be a process indicator directly linked with GPP. All models were able to capture the seasonal dynamics of GPP well, but underestimation of the growing season peaks were a common issue. The LRF was the only model tending to overestimate GPP. Estimation of interannual variability in cumulative GPP was less accurate than the single-year models and will need further development. In general, all models performed well on local scale and demonstrated their feasibility for upscaling GPP in northern forest ecosystems using Sentinel-2 data.

The heading date is an important fundamental trait in rice, which determines the length of growing duration and influences final yield. The traditional method to measure rice heading date involves frequent field work based on manual observations, which is slow, often subjective and feasible only in small areas. In this study, a Random Forest model was used to remotely estimate rice full heading (FH) date by unmanned aerial vehicle (UAV) imaging over the study sites throughout rice growing periods. The model using time-series Normalized Difference Vegetation Index (NDVI) and Normalized Difference Red Edge index (NDRE), retrieved from UAV multi-spectral images, was able to accurately estimate FH date for more than 1000 rice cultivars with root mean square errors below 4 days. The developed model was applied to map rice FH date variations under different environments. The results showed that most rice cultivars tend to heading later in response to colder temperatures while heading earlier at higher planting density, which has the sounded biological background. This study shows the great potential of using remote sensing method to assist in breeding studies, which is easy to implement across many fields and seasons, evaluating and comparing the crop trait for the large number of cultivars with high efficiency at low cost.

In 2002 and 2020–2022, KH-9 HEXAGON mapping camera system (MCS) and panoramic camera system (PCS) images were made available to the public, respectively. Although great efforts have been made by the scientific community to develop applications that utilize KH-9 HEXAGON images, little attention has been paid to de-noising and contrast enhancement of these images particularly over urban landscapes. This paper focuses on developing a de-noising and contrast enhancement pipeline for KH-9 HEXAGON MCS and PCS over urban regions. The proposed approach employs first a wavelet transform trained using a suite of ‘degree of over-smoothing’ metrics (DOSM) for image de-noising. These metrics are sensitive to structure, texture, edges and local homogeneity of image objects. Then the de-noised image is subjected to the multi-resolution Top-hat to optimize the contrast. This method incorporates a range of shapes and neighborhoods at multiple scales. The method was applied to a KH-9 HEXAGON MCS image (acquired in 1975) and PCS image (acquired in 1974) representing a complex urban landscape, to support comprehensive evaluation under a range of settings. Performance was assessed against three state-of-the-art benchmark approaches: residual learning (deep learning), blind deconvolution and spatial filtering. To evaluate the performance of the proposed pipeline against the benchmarks, we employed the saturation image edge difference standard-deviation, co-occurrence metrics and the semivariogram. Additionally, the potential applications of pre-processed results were demonstrated using change detection, identification reference points and stereo images. The proposed method not only improved the quality of the KH-9 image across the different urban landscape types, but also preserved the original spatial characteristics of the image in comparison with the benchmark methods. At a time when understanding the nature of our changing planet is paramount, the proposed pipeline should be of great benefit to investigators wishing to use KH program images to extend their historical or time-series analyses further back in time.

Unexploded munitions are some of the most enduring remnants of conflicts around the world. Their effects on the economy, health, environment, and post-conflict rehabilitation are long reaching and devastating for the areas they plague. With the advancements in very high spatial resolution (VHR) satellite multispectral imaging at sub-meter resolution, it becomes possible to detect object attributes at the scale of individual impacts (craters) of heavy weapon shelling. Manual identification and delineation of artillery craters in satellite imagery is time and resource consuming, especially when large territories and volumes of VHR data are considered. Therefore, automatic image processing methods should be explored. Here, we evaluate the application of a deep learning approach for identifying and mapping artillery craters in agricultural fields in Eastern Ukraine during the onset of armed conflict in 2014. The model was applied to pansharpened multispectral VHR imagery acquired by the WorldView-2 satellite at 0.5-m spatial resolution. The model can detect artillery craters with producer's accuracy (PA) (or recall) of 0.671 and user's accuracy (UA) (or precision) of 0.392 in terms of crater area and shape, and PA of 0.559 and UA of 0.427 in terms of binary crater identification. The model's performance is dependent on crater size. Reliability of crater detection and mapping improves as the size of craters increases. For example, for craters larger than 60 m² PA is 0.803 and UA is 0.449 (per-pixel), and PA is 0.891 and UA is 0.721 (per-object). Overall, the model prioritizes PA over UA, i.e., omission error over commission error, and is better at detecting craters than their shapes. We applied the trained model to a separate, 858 km² subregion of Donetsk oblast to automatically estimate and map the locations, number and area of artillery craters. Our estimates revealed over 22,000 craters in the subregion, which occupy an area of 1.2 km², or 0.14% of the region, primarily in agricultural fields. The availability of such crater maps is extremely valuable within demining and chemical decontamination efforts and can assist in assessing the impact of warfare on agriculture and the environment. We outline the current limitations of the proposed approach and avenues for further research for improving artillery crater detection and mapping.

Contemporary fire dynamics is one of the most complex and least understood land surface phenomena. Global fire controls related to climate, vegetation, and anthropogenic activity are usually intertwined, and difficult to disentangle in a quantitative way. Here, we leveraged an ensemble of five machine learning (ML) models and multiple satellite-based observations to conduct global fire modeling for three fire metrics (burned area, fire number, and fire size), and quantified driving mechanisms underlying annual fire changes in a spatially resolved manner for the period 2003–2019. Ensemble learning is a meta-approach that combines multiple ML predictions to improve accuracy, robustness, and generalization performance. We found that the optimized ensemble ML well reproduced annual dynamics of global burned area (R² = 0.90, P < 0.001), total fire numbers (R² = 0.86, P < 0.001), and averaged fire size (R² = 0.70, P < 0.001). Additionally, the ensemble ML captured key spatial patterns of multi-year mean magnitudes, annual variabilities, anomalies, and trends for different fire metrics. Our ML-based fire attributions further highlighted the dominant role of enhanced anthropogenic activity in reducing global burned area (−1.9 Mha/yr, P < 0.01), followed by climate control (−1.3 Mha/yr, P < 0.01) and insignificant positive vegetation control (0.4 Mha/yr, P = 0.60). Spatially, climate dominated a much larger burned area (53.7%) than human (23.4%) or vegetation control (22.9%); however, the counteracting effects from regional wetting and drying trends weakened the net climate impacts on global burned area. The fire number and fire size exhibited similar spatial control patterns with burned area; globally, however, fire number tended to be more affected by climate while fire size more influenced by human activities. Overall, our study confirmed the feasibility and efficiency of ensemble ML in global fire modeling and subsequent control attributions, providing a better understanding of contemporary fire regimes and contributing to robust fire projections in a changing environment.

Snow cover extent and distribution over the years have a significant impact on hydrological, terrestrial, and climatologic processes. Snow cover mapping accuracy using remote sensing data is then particularly important. This study analyses Landsat-8 NDSI snow cover datasets over time and space using different NDSI-based approach. The objectives are (i) to investigate the relation snow-NDSI with different environmental variables, (ii) to evaluate the accuracy of the common NDSI threshold of 0.4 against in-situ snow depth measurement and (iii) to develop a method that optimises snow cover mapping accuracy and minimises snow cover detection errors of omission and commission. Landsat-8 snow cover datasets were compared to ground snow depth measurements of climate stations over Switzerland for the period 2014–2020. It was found that there is a consistent relationship between NDSI values and land cover type, elevation, seasons, and snow depth measurements. The global NDSI threshold of 0.4 may not be always optimal for the Swiss territory and tends to underestimate the snow cover extent. Best NDSI thresholds vary spatially and are generally lower than 0.4 for the three snow depth threshold tested. We therefore propose a new spatiotemporal NDSI method to maximize snow cover mapping accuracy by using a generalized linear mixed model (GLMM). This model uses three environmental variables (i.e., elevation, land cover type and seasons) and raw NDSI values and improves snow cover mapping accuracy by 24% compared to the fixed threshold of 0.4. By using this method omissions errors decrease considerably while keeping a very low value of commission errors. This method will then be integrated in the Snow Observation from Space (SOfS) algorithm used for snow detection in Switzerland.