Pixel Time Series Research Articles

Trend of Changes in Phenological Components of Iran’s Vegetation Using Satellite Observations

Investigating vegetation changes, especially plant phenology, can yield valuable information about global warming and climate change. Time series satellite observations and remote sensing methods offer a great source of information on distinctions and changing aspects of vegetation. The current study aimed to determine the trend and rate of changes in some phenological components of Iran’s vegetation. In this regard, the current study employed the daily NDVI (Normalized Difference Vegetation Index) product of the AVHRR sensor with a spatial resolution of 0.05° × 0.05°, named AVH13C1. Then, using the HANTS algorithm, images of amplitude zero, annual amplitude, and annual phase were prepared annually from 1982 to 2019. Using TIMESAT software, the starting, end, and length of time of growing season were calculated for each pixel time series to prepare annual maps. The Mann–Kendall statistical test was used to investigate the significance of changes during the study period. On average in the entire area of Iran, the annual phase was declining with a trend of −0.6° per year, and the time for the start and end of the season was declining by −0.3 and −0.65 days per year, respectively. Major changes were noticed in the northeast, west, and northwest regions of Iran, where the annual phase declined with a trend of −0.9° per year. Since the annual growth cycle of the plant (equivalent to 356 days) was in the form of a sinusoidal signal, and the angular changes in the sine wave were between zero and 360°, each degree of change was equivalent to 1.01 days per year. Therefore, the reduction in the annual phase by −0.9 degrees almost means a change in the time (due to the earlier negative start phase) of the start of the annual growth signal by −0.9 days per year. The time of the start and end of the growing season declined by −0.6 and −1.33 days per year, respectively. The reduction in annual phase and differences in time of the starting season from 1982 to 2019 indicate the acceleration and earlier initiation of various phenological processes in the area.

ANALYSIS OF URBAN LAND USE CHANGE USING REMOTE SENSING AND DIFFERENT CHANGE DETECTION TECHNIQUES: THE CASE OF ANKARA PROVINCE

Abstract. This study aims to use remote sensing techniques to map the urban region of Ankara from the past to the present, assessing the nature, magnitude and direction of changes within the area, including the transformation of LULC classes and explaining the driving forces behind these transformations. The study encompasses three stages. Firstly, Landsat 7 ETM+ images from 2000 and Sentinel-2 satellite images from 2020 were obtained for Ankara city and surroundings through the Google Earth Engine (GEE) platform. Image classification was conducted for both 2000 and 2020 using 'Blue', 'Green', 'Red', 'Vegetation Red Edge1', 'Vegetation Red Edge2', 'Vegetation Red Edge3', 'NIR', 'Vegetation Red Edge4', 'Water vapour', ' SWIR1', 'SWIR2' bands, as well as 'NDWI', 'NDVI', 'NDBI' indices on the GEE platform. LULC was classified using the Random Forest (RF) classifier, which included six classes: urban area, forest, water surfaces, open areas, agricultural areas and roads. Secondly, the LULC maps of the 2000 and 2020 images were classified using RF. The study employed the 'Categorical Change, Pixel Value Change and Time Series Change' methods to determine the transformations between LULC categories. Specifically, the urban change within the study area increased by 70% between 2000 and 2020. Over the past 20 years, from 2000 to 2020, the urban areas in Ankara expanded by 170%. Consequently, accurately determining the nature, magnitude and direction of urban development using remote sensing data offers valuable baseline information for various disciplines related to spatial planning at local and national scales.

Automatically drawing vegetation classification maps using digital time‐lapse cameras in alpine ecosystems

AbstractAlpine ecosystems are particularly vulnerable to climate change. Monitoring the distribution of alpine vegetation is required to plan practical conservation activities. However, conventional field observations, airborne and satellite remote sensing are difficult in terms of coverage, cost and resolution in alpine areas. Ground‐based time‐lapse cameras have been used to observe the regions' snowmelt and vegetation phenology and offer significant advantages in terms of cost, resolution and frequency. However, they have not been used in research monitoring of vegetation distribution patterns. This study proposes a novel method for drawing georeferenced vegetation classification maps from ground‐based imagery of alpine regions. Our approach had two components: vegetation classification and georectification. The proposed vegetation classification method uses a pixel time series acquired from fall images, utilizing the fall leaf color patterns. We demonstrated that the performance of the vegetation classification could be improved using time‐lapse imagery and a Recurrent Neural Network. We also developed a novel method to accurately transform ground‐based images into georeferenced data. We propose the following approaches: (1) an automated procedure to acquire Ground Control Points and (2) a camera model that considers lens distortions for accurate georectification. We demonstrated that the proposed approach outperforms conventional methods, in addition to achieving sufficient accuracy to observe the vegetation distribution on a plant‐community scale. The evaluation revealed an F1 score and root‐mean‐square error of 0.937 and 3.4 m in the vegetation classification and georectification, respectively. Our results highlight the potential of inexpensive time‐lapse cameras to monitor the distribution of alpine vegetation. The proposed method can significantly contribute to the effective conservation planning of alpine ecosystems.

Multi-sensor detection of spring breakup phenology of Canada's lakes

The ice phenology of freshwater lakes throughout the Northern Hemisphere has undergone important climate-induced shifts over the past century. In Canada's North, where freshwater lakes and wetlands cover 15 to 40% of the landscape, monitoring ice phenology is vital to understand its impacts on climate, socio-economic, ecological, and hydrological systems. The rapid and dynamic nature of ice phenology events has restricted monitoring efforts to the use of satellite sensors with frequent revisit times (e.g., MODIS, AVHRR), but their low resolution (e.g., > 500 m) limits observations to larger water bodies. However, the increased abundance of high-resolution open-access satellite imagery combined with the rise of cloud-computing technologies has provided opportunities to reduce the trade-off between temporal (i.e., revisit time) and spatial (i.e., pixel size) resolution allowing for lake ice monitoring over broad scales. In this study, we present the Open Pixel-based Earth eNgine Ice (OPEN-ICE) algorithm implemented in Google Earth Engine (GEE), which classifies imagery from multiple open-access optical sensors, then combines them to construct dense annual time series of ice-water observations and estimate pixel spring breakup dates at a 30-m resolution. Using Landsat 7 ETM+, Landsat 8 OLI, and Sentinel-2 MSI scenes over lakes spanning northern latitudes, we build reference datasets to train decision trees that discriminate between ice, water, and clouds. We combine ice-water classifications from each sensor into annual time series and remove misclassifications with a temporal filter applied using a pixel-wise logistic regression. We then detect the sequence of transition from ice to water in each pixel's time series to estimate the occurrence of breakup each year. We deploy the OPEN-ICE algorithm over all freshwater pixels of Canada for the period of 2013 to 2021. Spring ice phenology events estimated by OPEN-ICE show high accuracy when compared to whole-lake breakup dates measured by the Canadian Ice Service in 105 lakes across 9 years, with mean bias errors of −1.10 and − 0.69 days for breakup start and end, respectively. We apply the OPEN-ICE algorithm to 4000 lakes across Canada and evaluate differences in breakup dates across ecozones and lake sizes. Our new OPEN-ICE tool provides accurate estimates of annual spring breakup events applicable across all boreal and arctic regions to monitor the rapid changes taking place in these vulnerable ecosystems.

Demonstration of large area land cover classification with a one dimensional convolutional neural network applied to single pixel temporal metric percentiles

Over large areas, land cover classification has conventionally been undertaken using satellite time series. Typically temporal metric percentiles derived from single pixel location time series have been used to take advantage of spectral differences among land cover classes over time and to minimize the impact of missing observations. Deep convolutional neural networks (CNNs) have demonstrated potential for land cover classification of single date images. However, over large areas and using time series their application is complicated because they are sensitive to missing observations and they may misclassify small and spatially fragmented surface features due to their spatial patch-based implementation. This study demonstrates, for the first time, a one-dimensional (1D) CNN single pixel time series land classification approach that uses temporal percentile metrics and that does not have these issues. This is demonstrated for all the Conterminous United States (CONUS) considering two different 1D CNN structures with 5 and 8 layers, respectively. CONUS 30 m land cover classifications were derived using all the available Landsat-5 and -7 imagery over a seven-month growing season in 2011 with 3.3 million 30 m land cover class labelled samples extracted from the contemporaneous CONUS National Land Cover Database (NLCD) 16 class land cover product. The 1D CNNs and, a conventional random forest model, were trained using 10%, 50% and 90% samples, and the classification accuracies were evaluated with an independent 10% proportion. Temporal metrics were classified using 5, 7 and 9 percentiles for each of five Landsat reflective wavelength bands and their eight band ratios. The CONUS and detailed 150 × 150 km classification results demonstrate that the approach is effective at scale and locally. The 1D CNN classification land cover class boundaries were preserved for small axis dimension features, such as roads and rivers, with no stripes or anomalous spatial patterns. The 8-layer 1D CNN provided the highest overall classification accuracies and both the 5-layer and 8-layer 1D CNN architectures provided higher accuracies than the random forest by 1.9% - 2.8% which as all the accuracies were > 83% is a meaningful increase. The CONUS overall classification accuracies increased marginally with the number of percentiles (86.21%, 86.40%, and 86.43% for 5, 7 and 9 percentiles, respectively) using the 8-layer 1D-CNN. Class specific producer and user accuracies were quantified, with lower accuracies for the developed land, crop and pasture/hay classes, but no systematic pattern among classes with respect to the number of temporal percentiles used. Application of the trained model to a different year of CONUS Landsat ARD showed moderately decreased accuracy (80.79% for 7 percentiles) that we illustrate is likely due to different intra-annual surface variations between years. These encouraging results are discussed with recommended research for deep learning using temporal metric percentiles.

LAND COVER DYNAMIC CHANGE MODES ANALYSIS BASED ON EMPIRICAL MODE DECOMPOSITION

Abstract. The dynamic change mode of land cover not only reflects the change characteristics of the land cover itself, but also reflects the influence exerted by external factors. Empirical mode decomposition is an effective time-frequency analysis method for nonlinear and nonstationary signals. In this paper, empirical mode decomposition firstly uses to each pixel time series to get its intrinsic mode functions, each intrinsic mode function relates to change mode with some time scale, clustering those intrinsic mode functions over the image, and different land cover change modes are obtained. The experimental dataset used is MODIS subset from the Jan 2011 to Dec 2020, located in the Ningxia Hui Autonomous Region, China. The clustering results of half-annual cycle and one-year cycle of intrinsic mode functions show the influence of local climate and policy factors on land cover change.

A Modification to Phase Estimation for Distributed Scatterers in InSAR Data Stacks

To improve the spatial density and quality of measurement points in multitemporal interferometric synthetic aperture radar, distributed scatterers (DSs) should be processed. An essential procedure in DS interferometry is phase estimation, which reconstructs a consistent phase series from all available interferograms. Influenced by the well-known suboptimality of coherence estimation, the performance of the state-of-the-art phase estimation algorithms is severely degraded. Previous research has addressed this problem by introducing the coherence bias correction technique. However, the precision of phase estimation is still insufficient because of the limited correction capabilities. In this paper, a modified phase estimation approach is proposed. Particularly, by incorporating the information on both interferometric coherence and the number of looks, a significant bias correction to each element of the coherence magnitude matrix is achieved. The bias-corrected coherence matrix is combined with advanced statistically homogeneous pixel selection and time series phase optimization algorithms to obtain the optimal phase series. Both the simulated and Sentinel-1 real data sets are used to demonstrate the superiority of this proposed approach over the traditional phase estimation algorithms. Specifically, the coherence bias can be corrected with considerable accuracy by the proposed scheme. The mean bias of coherence magnitude is reduced by more than 29%, and the standard deviation is reduced by more than 18% over the existing bias correction method. The proposed approach achieves higher accuracy than the current methods over the reconstructed phase series, including smoother interferometric phases and fewer outliers.

Comparison of common classification strategies for large-scale vegetation mapping over the Google Earth Engine platform

• We assessed the effect of various sample properties on classification performance. • The use of pixel time series from fusion data achieved optimal results. • Target class proportions and sample time frames affected accuracy. • Suboptimal methods had a larger impact on the accuracy of minority classes. • The combined use of Sentinel MSI and SAR data provided accurate vegetation maps. Vegetation resources have an essential role in sustainable development due to their close relationship with natural resource management and environmental protection. The monitoring of land use and cover is key for a more sustainable management of these resources, and Earth Observation satellites have provided an increasingly powerful platform for performing this task. To date, numerous classification algorithms have been developed for vegetation mapping, but a comparative evaluation of the strategies commonly used in large-scale applications (50000 km 2 and above) is lacking. We developed a classification framework based on random forests and Sentinel data within the Google Earth Engine platform to assess the performance of various strategies over a complex landscape with a wide range of vegetation classes, plot configurations, and agricultural practices. These strategies differed in key aspects related to the characteristics of the classifier and the sample, and were evaluated by class-specific and general accuracy metrics. We found that the use of pixel time series features from fusion data, class-balanced labels, and multi-season time frames enhanced overall performance by 1.3%–14.1% over alternative approaches, but in some cases also generated tradeoffs of 1.6%–8.5% between recall and precision. In general, suboptimal strategies were particularly ineffective for the detection of infrequent classes. Finally, the different parameter values used in the random forests did not have a significant influence over the results. Our results demonstrate the importance of algorithm design in the effective classification of multiple vegetation classes, corroborate the usefulness of Sentinel data to generate mapping products of high resolution and accuracy, and highlight the importance of cloud computing tools for the development of vegetation mapping tools in large-scale applications. These findings provide general guidelines for the design of future classification frameworks, which are necessary for facilitating a sustainable management of natural resources worldwide.

Integration of multiple earthquakes precursors before large earthquakes: A case study of 25 April 2015 in Nepal

In this work, we analyze magnetic data, Land Surface Temperature (LST), Outgoing Longwave Radiation (OLR), Air Temperature (AT) and Relative Humidity (RH) around the epicenter before and after the large EQs occurred on April 25, 2015 in Nepal. Daytime and nighttime LST values were retrieved from Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Terra satellite. We used the magnetic data acquired by Swarm satellites in the top side ionosphere, and daily OLR data was obtained from the United States National Oceanic and Atmospheric Administration (NOAA) polar-orbiting satellites. Moreover, AT and RH values were provided from the nearest meteorological ground station to the epicenter located at Katmandu airport. As data analysis, short-term (15 days before and 5 days after the earthquakes) and medium-term (2 months before and 1 month after the earthquakes) evaluations were carried out. For short-term analysis, the results of the magnetic field in the topside ionosphere detected an interesting anomaly on April 12, 2015. Besides, nighttime LST anomaly was observed on April 13, 2015, 14 days before the earthquakes, which was also an anomalous day for AT and RH analysis. Moreover, in the OLR analysis, April 12, 2015 was anomalous day based on the time series of the epicentral pixel that coincides with the magnetic field anomaly. Concerning the medium-term nighttime LST analysis, 13th April was the day, which may be associated with the earthquakes. For the medium-term OLR analysis, April 12, 2015 was found to be related to the seismic activity based on the epicentral pixel. Besides, the RH anomaly on 13 April was the only mutual day with the anomalies of LST and OLR for the medium-term analysis. Based on the short-term and medium-term results, these two days, 12 and April 13, 2015, may be associated with the preparation for the earthquake occurrence. Overall, it is concluded that these methods could be effective and reliable in detecting anomalies that precede upcoming EQs with further investigations.

Dynamic Cell Imaging: application to the diatom Phaeodactylum tricornutum under environmental stresses

ABSTRACT The dynamic movement of cell organelles is an important and poorly understood component of cellular organization and metabolism. In this work we present a non-invasive non-destructive method (Dynamic Cell Imaging, DCI) based on light scattering and interferometry to monitor dynamic events within photosynthetic cells using the diatom Phaeodactylum tricornutum as a model system. For this monitoring we acquire for a few seconds movies of the signals that are related to the motion of dynamic structures within the cell (denoted scatterers), followed by a statistical analysis of each pixel time series. Illuminating P. tricornutum with LEDs of different wavelengths associated with short pulsed or continuous-wave modes of illumination revealed that dynamic movements depend on chloroplast activity, in agreement with the reduction in the number of pixels with dynamic behaviour after addition of photosystem II inhibitors. We studied P. tricornutum under two environmentally relevant stresses, iron and phosphate deficiency. The major dynamic sites were located within lipid droplets and chloroplast envelope membranes. By comparing standard deviation and cumulative sum analyses of the time series, we showed that within the droplets two types of scatterer movement could be observed: random motion (Brownian type) but also anomalous movements corresponding to a drift which may relate to molecular fluxes within a cell. The method appears to be valuable for studying the effects of various environments on microalgae in the laboratory as well as in natural aquatic environments. HIGHLIGHTS Light scattering is an alternative to fluorescence to rapidly evidence dynamic processes. Lipid droplets are the major metabolic active sites under stress. A non-destructive visualization method suitable for laboratory microalgae and aquatic samples.

EVALUATION OF PIXEL SELECTION METHODS FOR TRAFFIC INFRASTRUCTURE MONITORING USING SENTINEL-1 INSAR

Abstract. The Synthetic Aperture Radar (SAR) satellite Sentinel-1 provides excellent traffic infrastructure monitoring capabilities due to its short revisit time, wide-scale coverage and free of charge data policy. However, pixels from Sentinel-1 have a medium spatial resolution so that traffic infrastructure is only covered by a few pixels. Moreover, Interferometric Synthetic Aperture Radar (InSAR) yields deformation time series for coherent pixels only, thus limiting the number of pixels reporting the ground motion at traffic infrastructure. Although various InSAR time series methods have been successfully applied for traffic infrastructure monitoring, the selection of appropriate methods achieving a high pixel density has not yet been evaluated. In this study, we test whether we can improve the monitoring capabilities by combining different InSAR time series methods. On the one hand, we applied widely used Stanford Method for Persistent Scatterers (StaMPS) as a baseline method to retrieve the deformation time series. On the other hand, we enhanced the phase quality in a pre-processing step using the Phase-Linking (PL) approach and afterwards estimated the deformation time series also with StaMPS based on pixels selected by PL. We compared and evaluated the achieved pixel densities from both methods at main roads, highways and railways in a study area in Germany. We found that the InSAR time series methods selected complementary sets of pixels at all traffic infrastructure types. Moreover, the results indicate a great potential for railway monitoring using Sentinel-1 InSAR due to both high pixel density and homogeneous spatial distribution of pixels. Interestingly, coherent scatterers selected by StaMPS were observed to coincide with large traffic signs at highways which show a double-bounce scattering in the amplitude images. This study shows the benefits of combining PL and StaMPS for increased pixel density at traffic infrastructure and confirms Sentinel-1 data as a suitable data source for traffic infrastructure monitoring.

Application of the artificial intelligence approach and remotely sensed imagery for soil moisture evaluation

AbstractThe current research attempts to present a modeling framework for determining soil moisture conditions by using remotely sensed imagery products. In this way, identifying various pixels with similar patterns from satellite images could be a reliable method to have an appropriate view over the soil moisture condition of a particular region. In this context, an artificial intelligence-based self-organizing map (SOM) method is employed to classify homogenous pixels over Phoenix, which is located in the south of Arizona, utilizing parameters extracted from satellite images. The central pixels of clusters are selected as the cluster indicator, with one from each cluster. Then, feed-forward neural networks (FFNNs) consisting of three layers of input, hidden, and output are trained by employing the extracted satellite images time series of the central pixels of the clusters. Finally, the soil moisture conditions of the representative pixels of the clusters are simulated by the trained models. The results reveal the suitability of SOM-based clustering to identify the specific points by which soil moisture can represent the soil moisture condition over the related regions. The proposed methodology and obtained results can be further used to provide a cost-effective method to determine the soil moisture condition of the region by reducing the costs of monitoring.HIGHLIGHTS An SOM is used to cluster homogenous pixels. The soil moisture conditions of the representative pixel for each cluster are simulated by using an ANN. The results reveal the suitability of the SOM clustering method to identify the specific points by which the soil moisture can represent the soil moisture condition.

10 m crop type mapping using Sentinel-2 reflectance and 30 m cropland data layer product

The 30 m resolution U.S. Department of Agriculture (USDA) crop data layer (CDL) is a widely used crop type map for agricultural management and assessment, environmental impact assessment, and food security. A finer resolution crop type map can potentially reduce errors related to crop area estimation, field size characterization, and precision agriculture activities that requires crop growth information at scales finer than crop field. This study is to develop a method for crop type mapping using Sentinel-2 10 m bands (i.e., red, green, blue, and near-infrared) and to examine the benefit of the derived 10 m crop type map. The crop type mapping was conducted for two study areas with significantly different field sizes and crop types in South Dakota and California, respectively. The Sentinel-2 10 m surface reflectance and the derived normalized difference vegetation index (NDVI) acquired in the 2019 growing season were used to generate monthly median composites as classification input. The training and evaluation samples were derived from CDL by (i) finding good quality 30 m CDL pixels and (ii) identifying a single representative Sentinel-2 10 m pixel time series for each 30 m good quality CDL pixel. The random forest algorithm was trained using 80% of the samples and evaluated using the 20% remaining samples, and the results showed high overall accuracies of 94% and 83% for South Dakota and California study areas, respectively. The major crops in both study areas obtained high user’s and producer’s accuracies (>87%). There is a good agreement between the class proportions in the 10 m crop type map and 30 m CDL for both study areas with R2 ≥ 0.94 and root mean square error (RMSE) ≤ 3%. More importantly, compared to the 30 m CDL, the 10 m crop type map has much less salt-pepper and crop boundary-aliasing effects and defines better the small surface features (e.g., small fields, roads, and rivers). The potential of the method for large area 10 m crop type mapping is discussed.

Seasonal Surface Fluctuation of a Slow-Moving Landslide Detected by Multitemporal Interferometry (MTI) on the Huafan University Campus, Northern Taiwan

A slow-moving landslide on the Huafan University campus, which is located on a dip slope in northern Taiwan, has been observed since 1990. However, reliable monitoring data are difficult to acquire after 2018 due to the lack of continuous maintenance of the field measurement equipment. In this study, the multitemporal interferometry (MTI) technique is applied with Sentinel-1 SAR images to monitor the slow-moving landslide from 2014–2019. The slow-moving areas detected by persistent scatterer (PS) pixels are consistent with the range of previous studies, which are based on in situ monitoring data and field surveys. According to the time series of the PS pixels, a long period gravity-induced deformation of the slow-moving landslide can be clearly observed. Moreover, a short period seasonal surface fluctuation of the slow-moving landslide, which has seldom been discussed before, can also be detected in this study. The seasonal surface fluctuation is in-phase with precipitation, which is inferred to be related to the geological and hydrological conditions of the study area. The MTI technique can compensate for the lack of surface displacement data, in this case, the Huafan University campus, and provide information for evaluating and monitoring slow-moving landslides for possible landslide early warning in the future.

Spatio-Temporal Mixed Pixel Analysis of Savanna Ecosystems: A Review

Reliable estimates of savanna vegetation constituents (i.e., woody and herbaceous vegetation) are essential as they are both responders and drivers of global change. The savanna is a highly heterogenous biome with high variability in land cover types while also being very dynamic at both temporal and spatial scales. To understand the spatial-temporal dynamics of savannas, using Earth Observation (EO) data for mixed-pixel analysis is crucial. Mixed pixel analysis provides detailed land cover data at a sub-pixel level which are essential for conservation purposes, understanding food supply for herbivores, quantifying environmental change, such as bush encroachment, and fuel availability essential for understanding fire dynamics, and for accurate estimation of savanna biomass. This review paper consulted 197 studies employing mixed-pixel analysis in savanna ecosystems. The review indicates that studies have so far attempted to resolve the savanna mixed-pixel issues by using mainly coarse resolution data, such as Terra-Aqua MODIS and AVHRR and medium resolution Landsat, to provide fractional cover data. Hence, there is a lack of spatio-temporal mixed-pixel analysis for savannas at high spatial resolutions. Methods used for mixed-pixel analysis include parametric and non-parametric methods which range from pixel-unmixing models, such as linear spectral mixture analysis (SMA), time series decomposition, empirical methods to link the green vegetation parameters with Vegetation Indices (VIs), and machine learning methods, such as regression trees (RT) and random forests (RF). Most studies were undertaken at local and regional scale, highlighting a research gap for savanna mixed pixel studies at national, continental, and global level. Parametric methods for modeling spatio-temporal mixed pixel analysis were preferred for coarse to medium resolution remote sensing data, while non-parametric methods were preferred for very high to high spatial resolution data. The review indicates a gap for long time series spatio-temporal mixed-pixel analysis of savannas using high resolution data at various scales. There is potential to harmonize the available low resolution EO data with new high-resolution sensors to provide long time series of the savanna mixed pixel, which, according to this review, is missing.

BENCHMARKING OF CONVOLUTIONAL NEURAL NETWORK APPROACHES FOR VEGETATION LAND COVER MAPPING

Abstract. Satellite Image Time Series (SITS) are becoming available at high spatial, spectral and temporal resolutions across the globe by the latest remote sensing sensors. These series of images can be highly valuable when exploited by classification systems to produce frequently updated and accurate land cover maps. The richness of spectral, spatial and temporal features in SITS is a promising source of data for developing better classification algorithms. However, machine learning methods such as Random Forests (RF), despite their fruitful application to SITS to produce land cover maps, are structurally unable to properly handle intertwined spatial, spectral and temporal dynamics without breaking the structure of the data. Therefore, the present work proposes a comparative study of various deep learning algorithms from the Convolutional Neural Network (CNN) family and evaluate their performance on SITS classification. They are compared to the processing chain coined iota2, developed by the CESBIO and based on a RF model. Experiments are carried out in an operational context using with sparse annotations from 290 labeled polygons. Less than 80 000 pixel time series belonging to 8 land cover classes from a year of Sentinel-2 monthly syntheses are used. Results show on a test set of 131 polygons that CNNs using 3D convolutions in space and time are more accurate than 1D temporal, stacked 2D and RF approaches. Best-performing models are CNNs using spatio-temporal features, namely 3D-CNN, 2D-CNN and SpatioTempCNN, a two-stream model using both 1D and 3D convolutions.

Mapping Winter Crops Using a Phenology Algorithm, Time-Series Sentinel-2 and Landsat-7/8 Images, and Google Earth Engine

With the increasing population and continuation of climate change, an adequate food supply is vital to economic development and social stability. Winter crops are important crop types in China. Changes in winter crops planting areas not only have a direct impact on China’s production and economy, but also potentially affects China’s food security. Therefore, it is necessary to obtain information on the planting of winter crops. In this study, we use the time series data of individual pixels, calculate the temporal statistics of spectral bands and the vegetation indices of optical data based on the phenological characteristics of specific vegetation or crops and record them in the time series data, and apply decision trees and rule-based algorithms to generate annual maps of winter crops. First, we constructed a dataset combining all the available images from Landsat 7/8 and Sentinel-2A/B. Second, we generated an annual map of land cover types to obtain the cropland mask in 2019. Third, we generated a time series of a single cropland pixel, and calculated the phenological indicators for classification by extracting the differences in phenological characteristics of different crops: these phenological indicators include SOS (start of season), SDP (start date of peak), EOS (end of season), GUS (green-up speed) and GSL (growing-season length). Finally, we identified winter crops in 2019 based on their phenological characteristics. The main advantages of the phenology-based algorithm proposed in this study include: (1) Combining multiple sensor data to construct a high spatiotemporal resolution image collection. (2) By analyzing the whole growth season of winter crops, the planting area of winter crops can be extracted more accurately, and (3) the phenological indicators of different periods are extracted, which is conducive to monitoring winter crop planting information and seasonal dynamics. The results show that the algorithm constructed in this study can accurately extract the planting area of winter crops, with user, producer, overall accuracies and Kappa coefficients of 96.61%, 94.13%, 94.56% and 0.89, respectively, indicating that the phenology-based algorithm is reliable for large area crop classification. This research will provide a point of reference for crop area extraction and monitoring.

Google Earth Engine Sentinel-3 OLCI Level-1 Dataset Deviates from the Original Data: Causes and Consequences

In this study, we demonstrate that the Google Earth Engine (GEE) dataset of Sentinel-3 Ocean and Land Color Instrument (OLCI) level-1 deviates from the original Copernicus Open Access Data Hub Service (DHUS) data by 10–20 W m−2 sr−1μμm−1 per pixel per band. We compared GEE and DHUS single pixel time series for the period from April 2016 to September 2020 and identified two sources of this discrepancy: the ground pixel position and reprojection. The ground pixel position of OLCI product can be determined in two ways: from geo-coordinates (DHUS) or from tie-point coordinates (GEE). We recommend using geo-coordinates for pixel extraction from the original data. When the Sentinel Application Platform (SNAP) Pixel Extraction Tool is used, an additional distance check has to be conducted to exclude pixels that lay further than 212 m from the point of interest. Even geo-coordinates-based pixel extraction requires the homogeneity of the target area at a 700 m diameter (49 ha) footprint (double of the pixel resolution). The GEE OLCI dataset can be safely used if the homogeneity assumption holds at 2700 m diameter (9-by-9 OLCI pixels) or if the uncertainty in the radiance of 10% is not critical for the application. Further analysis showed that the scaling factors reported in the GEE dataset description must not be used. Finally, observation geometry and meteorological data are not present in the GEE OLCI dataset, but they are crucial for most applications. Therefore, we propose to calculate angles and extraterrestrial solar fluxes and to use an alternative data source—the Copernicus Atmosphere Monitoring Service (CAMS) dataset—for meteodata.

Joint Supervised Classification and Reconstruction of Irregularly Sampled Satellite Image Times Series

Recent satellite missions have led to a huge amount of Earth observation data, most of them being freely available. In such a context, satellite image time series have been used to study land use and land cover information. However, optical time series, such as Sentinel-2 or Landsat ones, are provided with an irregular time sampling for different spatial locations, and images may contain clouds and shadows. Thus, preprocessing techniques are usually required to properly classify such data. The proposed approach is able to deal with irregular temporal sampling and missing data directly in the classification process. It is based on Gaussian processes and allows to perform jointly the classification of the pixel labels as well as the reconstruction of the pixel time series. The method complexity scales linearly with the number of pixels, making it amenable in large-scale scenarios. Experimental classification and reconstruction results show that the method does not compete yet with state-of-the-art classifiers but yields reconstructions that are robust with respect to the presence of undetected clouds or shadows and does not require any temporal preprocessing.

Automatically selecting hot and cold pixels for satellite actual evapotranspiration estimation under different topographic and climatic conditions

The persistent monitoring of evapotranspiration (ET) over the regions suffering from water scarcity is critical for sustainable agricultural water management. Remote sensing provides time- and cost-effective capability to investigate daily ET rates at large scales. Satellite-based actual evapotranspiration (ETa) algorithms typically rely on specifying the upper and lower boundaries of ETa rate over agricultural and pasture fields, commonly known as hot (dry) and cold (wet) pixels selection. These boundaries are to be recognized by an expert through a subjective and labor-intensive task. In this study, a method has been introduced to automatically select appropriate anchor pixels (i.e., hot and cold pixels) independent from land use/cover maps with the simplest possible way, quickly applied even by an inexperienced operator. Subsequently, ETa was calculated using Mapping EvapoTranspiration at high Resolution with Internalized Calibration (METRIC), Surface Energy Balance Algorithm for Land (SEBAL), and Surface Energy Balance System (SEBS) algorithms and evaluated against measured data. In this method, the mountains and foothills were removed using the Shuttle Radar Topography Mission Digital Elevation Model (SRTM DEM) and the subsequent product was a slope mask. Then, filters were applied based on the Normalized Difference Vegetation Index (NDVI), Albedo, and Land Surface Temperature (LST) images to identify potential candidate pixels for hot and cold pixels. In the end, the best-conditioned pixel being closest to the meteorological station was selected. The method was assessed in five different regions with different topographic and climatic conditions. The selected pixels were first visually validated in Landsat images, and then the fluctuations and values were discussed in time series of anchor pixels and LST histograms. The visual interpretation was indicative of selecting the anchor pixels in fallow (hot pixel) and densely vegetated (cold pixel) surfaces. Also, the hot and cold pixels were suitably situated in the upper and lower quartiles of the LST histogram, respectively. The range of cold pixels variations throughout the study periods was lower compared with the hot pixels (44.2, 55.3, 35.5, 66.5, and 25.2 K for hot pixels against 34.2, 45.7, 25.6, 52.2, and 17 K for cold pixels) as expected, which emanated from the lower fluctuations of temperature over vegetation against the soil. The results were indicative of the better performance of METRIC compared with SEBAL and SEBS with greater values of R2 in all the regions. Therefore, using the introduced method, the expert subjective interference was eliminated and processing time reduced significantly from about 1 h per image to a few minutes.