Machine Learning Classification of Mediterranean Forest Habitats in Google Earth Engine Based on Seasonal Sentinel-2 Time-Series and Input Image Composition Optimisation

The sustainable management of natural heritage is presently considered a global strategic issue. Owing to the ever-growing availability of free data and software, remote sensing (RS) techniques have been primarily used to map, analyse, and monitor natural resources for conservation purposes. The need to adopt multi-scale and multi-temporal approaches to detect different phenological aspects of different vegetation types and species has also emerged. The time-series composite image approach allows for capturing much of the spectral variability, but presents some criticalities (e.g., time-consuming research, downloading data, and the required storage space). To overcome these issues, the Google Earth engine (GEE) has been proposed, a free cloud-based computational platform that allows users to access and process remotely sensed data at petabyte scales. The application was tested in a natural protected area in Calabria (South Italy), which is particularly representative of the Mediterranean mountain forest environment. In the research, random forest (RF), support vector machine (SVM), and classification and regression tree (CART) algorithms were used to perform supervised pixel-based classification based on the use of Sentinel-2 images. A process to select the best input image (seasonal composition strategies, statistical operators, band composition, and derived vegetation indices (VIs) information) for classification was implemented. A set of accuracy indicators, including overall accuracy (OA) and multi-class F-score (Fm), were computed to assess the results of the different classifications. GEE proved to be a reliable and powerful tool for the classification process. The best results (OA = 0.88 and Fm = 0.88) were achieved using RF with the summer image composite, adding three VIs (NDVI, EVI, and NBR) to the Sentinel-2 bands. SVM and RF produced OAs of 0.83 and 0.80, respectively.

Download Full-text

GOOGLE EARTH ENGINE: APPLICATION OF ALGORITHMS FOR REMOTE SENSING OF CROPS IN TUSCANY (ITALY)

ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences ◽

10.5194/isprs-archives-xlii-3-w12-2020-291-2020 ◽

2020 ◽

Vol XLII-3/W12-2020 ◽

pp. 291-296

Author(s):

J. P. Clemente ◽

G. Fontanelli ◽

G. G. Ovando ◽

Y. L. B. Roa ◽

A. Lapini ◽

...

Keyword(s):

Remote Sensing ◽

Time Series ◽

Google Earth ◽

Optical Data ◽

Support Vector ◽

Use Of Time ◽

Crop Types ◽

Google Earth Engine ◽

Agricultural Areas ◽

Sentinel 2

Abstract. Remote sensing has become an important mean to assess crop areas, specially for the identification of crop types. Google Earth Engine (GEE) is a free platform that provides a large number of satellite images from different constellations. Moreover, GEE provides pixel-based classifiers, which are used for mapping agricultural areas. The objective of this work is to evaluate the performance of different classification algorithms such as Minimum Distance (MD), Random Forest (RF), Support Vector Machine (SVM), Classification and Regression Trees (CART) and Na¨ıve Bayes (NB) on an agricultural area in Tuscany (Italy). Four different scenarios were implemented in GEE combining different information such as optical and Synthetic Aperture Radar (SAR) data, indices and time series. Among the five classifiers used the best performers were RF and SVM. Integrating Sentinel-1 (S1) and Sentinel-2 (S2) slightly improves the classification in comparison to the only S2 image classifications. The use of time series substantially improves supervised classifications. The analysis carried out so far lays the foundation for the integration of time series of SAR and optical data.

Download Full-text

An updated national-scale mangrove forest map in China Using Sentinel-1 and Sentinel-2 Time-Series Data with Google Earth Engine

10.5194/egusphere-egu21-5310 ◽

2021 ◽

Author(s):

Luojia Hu ◽

Wei Yao ◽

Zhitong Yu ◽

Yan Huang

Keyword(s):

Time Series ◽

High Resolution ◽

Mangrove Forest ◽

Google Earth ◽

Series Data ◽

National Scale ◽

Mangrove Area ◽

Google Earth Engine ◽

Forest Maps ◽

Sentinel 2

A high resolution mangrove map (e.g., 10-m), which can identify mangrove patches with small size (< 1 ha), is a central component to quantify ecosystem functions and help government take effective steps to protect mangroves, because the increasing small mangrove patches, due to artificial destruction and plantation of new mangrove trees, are vulnerable to climate change and sea level rise, and important for estimating mangrove habitat connectivity with adjacent coastal ecosystems as well as reducing the uncertainty of carbon storage estimation. However, latest national scale mangrove forest maps mainly derived from Landsat imagery with 30-m resolution are relatively coarse to accurately characterize the distribution of mangrove forests, especially those of small size (area < 1 ha). Sentinel imagery with 10-m resolution provide the opportunity for identifying these small mangrove patches and generating high-resolution mangrove forest maps. Here, we used spectral/backscatter-temporal variability metrics (quantiles) derived from Sentinel-1 SAR (Synthetic Aperture Radar) and sentinel-2 MSI (Multispectral Instrument) time-series imagery as input features for random forest to classify mangroves in China. We found that Sentinel-2 imagery is more effective than Sentinel-1 in mangrove extraction, and a combination of SAR and MSI imagery can get a better accuracy (F1-score of 0.94) than using them separately (F1-score of 0.88 using Sentinel-1 only and 0.895 using Sentinel-2 only). The 10-m mangrove map derived by combining SAR and MSI data identified 20,003 ha mangroves in China and the areas of small mangrove patches (< 1 ha) was 1741 ha, occupying 8.7% of the whole mangrove area. The largest area (819 ha) of small mangrove patches is located in Guangdong Province, and in Fujian the percentage of small mangrove patches in total mangrove area is the highest (11.4%). A comparison with existing 30-m mangrove products showed noticeable disagreement, indicating the necessity for generating mangrove extent product with 10-m resolution. This study demonstrates the significant potential of using Sentinel-1 and Sentinel-2 images to produce an accurate and high-resolution mangrove forest map with Google Earth Engine (GEE). The mangrove forest maps are expected to provide critical information to conservation managers, scientists, and other stakeholders in monitoring the dynamics of mangrove forest.

Download Full-text

Monitoring of water table dynamics in peatlands with OPTRAM: Towards globally applicable algorithms in Google Earth Engine using Landsat and Sentinel-2

10.5194/egusphere-egu21-4698 ◽

2021 ◽

Author(s):

Iuliia Burdun ◽

Michel Bechtold ◽

Viacheslav Komisarenko ◽

Annalea Lohila ◽

Elyn Humphreys ◽

...

Keyword(s):

Time Series ◽

Water Table ◽

Input Data ◽

Short Wave ◽

Google Earth ◽

Google Earth Engine ◽

Short Wave Infrared ◽

Northern Peatlands ◽

Sentinel 2

Fluctuations of water table depth (WTD) affect many processes in peatlands, such as vegetation development and emissions of greenhouse gases. Here, we present the OPtical TRApezoid Model (OPTRAM) as a new method for satellite-based monitoring of the temporal variation of WTD in peatlands. OPTRAM is based on the response of short-wave infrared reflectance to the vegetation water status. For five northern peatlands with long-term in-situ WTD records, and with diverse vegetation cover and hydrological regimes, we generate a suite of OPTRAM index time series using (a) different procedures to parametrise OPTRAM (peatland-specific manual vs. globally applicable automatic parametrisation in Google Earth Engine), and (b) different satellite input data (Landsat vs. Sentinel-2). The results based on the manual parametrisation of OPTRAM indicate a high correlation with in-situ WTD time-series for pixels with most suitable vegetation for OPTRAM application (mean Pearson correlation of 0.7 across sites), and we will present the performance differences when moving from a manual to an automatic procedure. Furthermore, for the overlap period of Landsat and Sentinel-2, which have different ranges and widths of short-wave infrared bands used for OPTRAM calculation, the impact of the satellite input data to OPTRAM will be analysed. Eventually, the challenge of merging different satellite missions in the derivation of OPTRAM time series will be explored as an important step towards a global application of OPTRAM for the monitoring of WTD dynamics in northern peatlands.

Download Full-text

Advancing the Mapping of Mangrove Forests at National-Scale Using Sentinel-1 and Sentinel-2 Time-Series Data with Google Earth Engine: A Case Study in China

Remote Sensing ◽

10.3390/rs12193120 ◽

2020 ◽

Vol 12 (19) ◽

pp. 3120

Author(s):

Luojia Hu ◽

Nan Xu ◽

Jian Liang ◽

Zhichao Li ◽

Luzhen Chen ◽

...

Keyword(s):

Time Series ◽

High Resolution ◽

Mangrove Forest ◽

Time Series Data ◽

Landsat Imagery ◽

Google Earth ◽

Series Data ◽

National Scale ◽

Google Earth Engine ◽

Sentinel 2

A high resolution mangrove map (e.g., 10-m), including mangrove patches with small size, is urgently needed for mangrove protection and ecosystem function estimation, because more small mangrove patches have disappeared with influence of human disturbance and sea-level rise. However, recent national-scale mangrove forest maps are mainly derived from 30-m Landsat imagery, and their spatial resolution is relatively coarse to accurately characterize the extent of mangroves, especially those with small size. Now, Sentinel imagery with 10-m resolution provides an opportunity for generating high-resolution mangrove maps containing these small mangrove patches. Here, we used spectral/backscatter-temporal variability metrics (quantiles) derived from Sentinel-1 SAR (Synthetic Aperture Radar) and/or Sentinel-2 MSI (Multispectral Instrument) time-series imagery as input features of random forest to classify mangroves in China. We found that Sentinel-2 (F1-Score of 0.895) is more effective than Sentinel-1 (F1-score of 0.88) in mangrove extraction, and a combination of SAR and MSI imagery can get the best accuracy (F1-score of 0.94). The 10-m mangrove map was derived by combining SAR and MSI data, which identified 20003 ha mangroves in China, and the area of small mangrove patches (<1 ha) is 1741 ha, occupying 8.7% of the whole mangrove area. At the province level, Guangdong has the largest area (819 ha) of small mangrove patches, and in Fujian, the percentage of small mangrove patches is the highest (11.4%). A comparison with existing 30-m mangrove products showed noticeable disagreement, indicating the necessity for generating mangrove extent product with 10-m resolution. This study demonstrates the significant potential of using Sentinel-1 and Sentinel-2 images to produce an accurate and high-resolution mangrove forest map with Google Earth Engine (GEE). The mangrove forest map is expected to provide critical information to conservation managers, scientists, and other stakeholders in monitoring the dynamics of the mangrove forest.

Download Full-text

Incorporating the Plant Phenological Trajectory into Mangrove Species Mapping with Dense Time Series Sentinel-2 Imagery and the Google Earth Engine Platform

Remote Sensing ◽

10.3390/rs11212479 ◽

2019 ◽

Vol 11 (21) ◽

pp. 2479 ◽

Cited By ~ 5

Author(s):

Huiying Li ◽

Mingming Jia ◽

Rong Zhang ◽

Yongxing Ren ◽

Xin Wen

Keyword(s):

Time Series ◽

Vegetation Index ◽

Google Earth ◽

Late Winter ◽

Mangrove Forests ◽

Mangrove Species ◽

Ndvi Time Series ◽

Google Earth Engine ◽

Species Mapping ◽

Sentinel 2

Information on mangrove species composition and distribution is key to studying functions of mangrove ecosystems and securing sustainable mangrove conservation. Even though remote sensing technology is developing rapidly currently, mapping mangrove forests at the species level based on freely accessible images is still a great challenge. This study built a Sentinel-2 normalized difference vegetation index (NDVI) time series (from 2017-01-01 to 2018-12-31) to represent phenological trajectories of mangrove species and then demonstrated the feasibility of phenology-based mangrove species classification using the random forest algorithm in the Google Earth Engine platform. It was found that (i) in Zhangjiang estuary, the phenological trajectories (NDVI time series) of different mangrove species have great differences; (ii) the overall accuracy and Kappa confidence of the classification map is 84% and 0.84, respectively; and (iii) Months in late winter and early spring play critical roles in mangrove species mapping. This is the first study to use phonological signatures in discriminating mangrove species. The methodology presented can be used as a practical guideline for the mapping of mangrove or other vegetation species in other regions. However, future work should pay attention to various phenological trajectories of mangrove species in different locations.

Download Full-text

Towards Global-Scale Seagrass Mapping and Monitoring Using Sentinel-2 on Google Earth Engine: The Case Study of the Aegean and Ionian Seas

Remote Sensing ◽

10.3390/rs10081227 ◽

2018 ◽

Vol 10 (8) ◽

pp. 1227 ◽

Cited By ~ 37

Author(s):

Dimosthenis Traganos ◽

Bharat Aggarwal ◽

Dimitris Poursanidis ◽

Konstantinos Topouzelis ◽

Nektarios Chrysoulakis ◽

...

Keyword(s):

Large Scale ◽

Anthropogenic Impacts ◽

Global Scale ◽

Google Earth ◽

Support Vector ◽

Learning Approaches ◽

Data Set ◽

Independent Field ◽

Google Earth Engine ◽

Sentinel 2

Seagrasses are traversing the epoch of intense anthropogenic impacts that significantly decrease their coverage and invaluable ecosystem services, necessitating accurate and adaptable, global-scale mapping and monitoring solutions. Here, we combine the cloud computing power of Google Earth Engine with the freely available Copernicus Sentinel-2 multispectral image archive, image composition, and machine learning approaches to develop a methodological workflow for large-scale, high spatiotemporal mapping and monitoring of seagrass habitats. The present workflow can be easily tuned to space, time and data input; here, we show its potential, mapping 2510.1 km2 of P. oceanica seagrasses in an area of 40,951 km2 between 0 and 40 m of depth in the Aegean and Ionian Seas (Greek territorial waters) after applying support vector machines to a composite of 1045 Sentinel-2 tiles at 10-m resolution. The overall accuracy of P. oceanica seagrass habitats features an overall accuracy of 72% following validation by an independent field data set to reduce bias. We envision that the introduced flexible, time- and cost-efficient cloud-based chain will provide the crucial seasonal to interannual baseline mapping and monitoring of seagrass ecosystems in global scale, resolving gain and loss trends and assisting coastal conservation, management planning, and ultimately climate change mitigation.

Download Full-text

Rapid, robust, and automated mapping of tidal flats in China using time series Sentinel-2 images and Google Earth Engine

Remote Sensing of Environment ◽

10.1016/j.rse.2021.112285 ◽

2021 ◽

Vol 255 ◽

pp. 112285 ◽

Cited By ~ 1

Author(s):

Mingming Jia ◽

Zongming Wang ◽

Dehua Mao ◽

Chunying Ren ◽

Chao Wang ◽

...

Keyword(s):

Time Series ◽

Tidal Flats ◽

Google Earth ◽

Automated Mapping ◽

Google Earth Engine ◽

Sentinel 2

Download Full-text

Testing the efficacy of hyperspectral (AVIRIS-NG), multispectral (Sentinel-2) and radar (Sentinel-1) remote sensing images to detect native and invasive non-native trees

10.1101/2021.01.01.425059 ◽

2021 ◽

Author(s):

M Arasumani ◽

Aditya Singh ◽

Milind Bunyan ◽

V.V. Robin

Keyword(s):

Remote Sensing ◽

Native Species ◽

Ground Truth ◽

Google Earth ◽

Support Vector ◽

Svm Classification ◽

Machine Learning Classification ◽

Tropical Grasslands ◽

Exotic Tree ◽

Sentinel 2

AbstractInvasive alien species (IAS) threaten tropical grasslands and native biodiversity and impact ecosystem service delivery, ecosystem function, and associated human livelihoods. Tropical grasslands have been dramatically and disproportionately lost to invasion by trees. The invasion continues to move rapidly into the remaining fragmented grasslands impacting various native grassland-dependent species and water streamflow in tropical montane habitats. The Shola Sky Islands of the Western Ghats host a mosaic of native grasslands and forests; of which the grasslands have been lost to exotic tree invasion (Acacias, Eucalyptus and Pines) since the 1950s. The invasion intensities, however, differ between these species wherein Acacia mearnsii and Pinus patula are highly invasive in contrast to Eucalyptus globulus. These disparities necessitate distinguishing these species for effective grassland restoration. Further, these invasive alien trees are highly intermixed with native species, thus requiring high discrimination abilities to native species apart from the non-native species.Here we assess the accuracy of various satellite and airborne remote sensing sensors and machine learning classification algorithms to identify the spatial extent of native habitats and invasive trees. Specifically, we test Sentinel-1 SAR and Sentinel-2 multispectral data and assess high spatial and spectral resolution AVIRIS-NG imagery identifying invasive species across this landscape. Sensor combinations thus include hyperspectral, multispectral and radar data and present tradeoffs in associated costs and ease of procurement. Classification methods tested include Support Vector Machine (SVM), Classification and Regression Trees (CART) and Random Forest (RF) algorithms implemented on the Google Earth Engine platform. Results indicate that AVIRIS-NG data in combination with SVM recover the highest classification skill (Overall −98%, Kappa-0.98); while CART and RF yielded < 90% accuracy. Fused Sentinel-1 and Sentinel-2 produce 91% accuracy, while Sentinel-2 alone yielded 91% accuracy with RF and SVM classification; but only with higher coverage of ground control points. AVIRIS-NG imagery was able to accurately (97%) demarcate the Acacia invasion front while Sentinel-1 and Sentinel-2 data failed. Our results suggest that Sentinel-2 images could be useful for detecting the native and non-native forests with more ground truth points, but hyperspectral data (AVIRIS-NG) permits distinguishing, native and non-native tree species and recent invasions with high precision using limited ground truth points. We suspect that large areas will have to be mapped and assessed in the coming years by conservation managers, NGOs to plan restoration, or to assess the success of restoration activities, and several data procurement and analysis steps may have to be simplified.

Download Full-text

Mapping Three Decades of Changes in the Brazilian Savanna Native Vegetation Using Landsat Data Processed in the Google Earth Engine Platform

Remote Sensing ◽

10.3390/rs12060924 ◽

2020 ◽

Vol 12 (6) ◽

pp. 924 ◽

Cited By ~ 9

Author(s):

Ane Alencar ◽

Julia Z. Shimbo ◽

Felipe Lenti ◽

Camila Balzani Marques ◽

Bárbara Zimbres ◽

...

Keyword(s):

Time Series ◽

Average Rate ◽

Landsat Imagery ◽

Google Earth ◽

Native Vegetation ◽

Brazilian Cerrado ◽

Statistical Decision ◽

Machine Learning Classification ◽

Cerrado Vegetation ◽

Google Earth Engine

Widespread in the subtropics and tropics of the Southern Hemisphere, savannas are highly heterogeneous and seasonal natural vegetation types, which makes change detection (natural vs. anthropogenic) a challenging task. The Brazilian Cerrado represents the largest savanna in South America, and the most threatened biome in Brazil owing to agricultural expansion. To assess the native Cerrado vegetation (NV) areas most susceptible to natural and anthropogenic change over time, we classified 33 years (1985–2017) of Landsat imagery available in the Google Earth Engine (GEE) platform. The classification strategy used combined empirical and statistical decision trees to generate reference maps for machine learning classification and a novel annual dataset of the predominant Cerrado NV types (forest, savanna, and grassland). We obtained annual NV maps with an average overall accuracy ranging from 87% (at level 1 NV classification) to 71% over the time series, distinguishing the three main NV types. This time series was then used to generate probability maps for each NV class. The native vegetation in the Cerrado biome declined at an average rate of 0.5% per year (748,687 ha yr−1), mostly affecting forests and savannas. From 1985 to 2017, 24.7 million hectares of NV were lost, and now only 55% of the NV original distribution remains. Of the remnant NV in 2017 (112.6 million hectares), 65% has been stable over the years, while 12% changed among NV types, and 23% was converted to other land uses but is now in some level of secondary NV. Our results were fundamental in indicating areas with higher rates of change in a long time series in the Brazilian Cerrado and to highlight the challenges of mapping distinct NV types in a highly seasonal and heterogeneous savanna biome.

Download Full-text

Landslides in Central Italy identified from Sentinel 2 multispectral imaging time series analysis with Google Earth Engine

10.5194/egusphere-egu2020-18214 ◽

2020 ◽

Author(s):

Marco Bartola ◽

Carla Braitenberg ◽

Carlo Bisci

Keyword(s):

Time Series ◽

Time Series Analysis ◽

Multispectral Imaging ◽

Spectral Response ◽

Central Italy ◽

Slope Angle ◽

Google Earth ◽

Series Analysis ◽

Google Earth Engine ◽

Sentinel 2

In 2016, Central Italy was hit by a months-lasting earthquake sequence that started off in August 24th 2016 with a Mw 6.2 earthquake which provoked severe damage to the towns of Accumoli (RI) and Amatrice (RI). The following October 30th 2016 earthquake (Mw 6.5), with epicenter in Norcia (PG) about 20 km NW of the first shock, triggered landslides in the area of Visso (MC), as reported by local newspapers.The purpose of this work is to individuate the areas affected by such landslides using the radiance variation recorded by multispectral images acquired by Sentinel 2. The time series analysis of the images has been carried out in Google Earth Engine environment, that allows access to the entire suite of available images. Due to the steep terrain, the shadowing effect of the hills was taken into account and comparison of images have been made only for those taken in the same seasonal moment of different years, thus guaranteeing the same solar elevation.It was found that the band of red was instrumental in identifying landslides along slopes made up of limestone, which is the typical outcrop of the area. Due to the extended time period between the images (July 2015 and July 2017), anthropogenic changes in land-use were present and had to be distinguished from landslides. A criterion involving the slope angle was developed, maintaining only the changes that had occurred on slopes steeper than 25&#176;, since man-made interventions giving similar spectral response are hardly done in steep areas. The slope analysis and correlation study with the extension and location of landslides was carried out using a Geographic Information System. (ESRI ArcGIS 10.5) The total extent of the area affected by the surveyed landslides is very large, having&#160; been estimated to be more than 200 000 m2.

Download Full-text