Climatic changes, catastrophic storm events, bark beetle calamities and unsustainable forest management pose a growing threat to our Alpine forests. Forest administrations and forest owners undertake great planning efforts and mitigation measures to prevent forest degradation and to maximize the economic potential of our forests. To do so, these forest entities depend on the availability of relevant and timely information on the status and dynamics of the Alpine forests.

Within the research project AlpMon (https://www.joanneum.at/digital/referenzprojekte/alpmon), a consortium of Joanneum Research, Catalysts and the EODC is developing forest monitoring tools to assist forest administration and forest owners. The AlpMon services transform Copernicus data into reliable information on up-to-date forest status that forest owners and forest administration require to better manage and protect our Alpine forests. Our innovative processing lines are based on time series analysis, high performance computing and web-based software solutions for an easy user interaction. Forest owners and forest administrations are enrolled as prototype users during service development to test and improve the services in operational forest management.

Two services are being offered by AlpMon: A near real-time forest alert system based on time series analysis to rapidly and continuously map and categorize forest changes in the Alpine region at 10m spatial resolution, and a forest map service that provides improved and updated maps on selected forest parameters (e.g. tree species, forest type, tree cover density, past forest changes). The near real-time forest change detection allows for a rapid assessment of storm damages, clear cuts and insect infestations, such as bark beetle attacks, and enables the initiation of rapid and effective countermeasures. We support the integration of 3D data, e.g. from airborne laser scanning, to rapidly estimate the affected timber volume after storm events. Forest statistics, such as affected area, timber volume or stand properties, can be calculated automatically and quickly for user defined areas of interest and used for administrative reporting duties. Both services are run in a cloud processing environment and could easily be deployed on other cloud computing systems (e.g. DIAS) as their design is independent of platform and operating systems. The users can comfortably access the services through a conventional web browser. The web interface also includes latest open data on land use and land cover from the Copernicus program.

We present the methodologies used in our near real-time forest change detection and show prototype examples of our services from the Alpine region. Near real-time forest monitoring is based on time series of Sentinel-2 data and an initial forest mask derived from airborne laser scanning or Sentinel-2 data. Time series stacks are pre-processed including atmospheric correction, topographic correction, improved cloud and snow masking and calculation of Tasselled Cap transformations. A historic harmonic model at single pixel level is derived from the pre-processed time series. The harmonic model parameters and the new observations are used as input to a Kalman filter which can better account for phenologic differences among individual years (e.g. later bud break in spring, earlier fall foliage) than the fixed harmonic model. Changes are flagged by calculating a statistical deviation between the individual new observation and the Kalman filtered model. If changes are detected in consecutive new observations, these changes are confirmed on single pixel level. Changes are then analysed and categorized based on change intensity and direction, spectral properties, neighbourhood rules and geometric constraints. The target categories for forest change are: clear cuts, forest thinnings, storm damages and insect infestations. The results are presented in a user web interface and can be downloaded as shapefiles or in geotiff format. Additionally, an alert function can be configured by the user for pre-defined areas of interest. If changes of a certain minimum extent or category are detected within this area, the user receives an alert notification by email.

We will show examples and validation results for prototype sites in Austria where the service is currently implemented and tested at operational stage. Examples will focus on the detection of storm damages and bark beetle infestations and we will present validation results for area based accuracy and timeliness of information delivery. The latter is an important issue to forest administrations, as they are bound by legislation to specific time frames for reporting forest damages. Also, the early detection of insect infestations can help to initiate forest management activities at an early stage and to prevent further outbreaks.

Earth Observation data has proven impactful to the development of sustainable forest management practices and conservation policies. At a time of the global trade and climate change, invasive biotic agents can create serious ecological and societal challenges requiring innovative solutions and near-real time monitoring. Detecting and characterizing changes, both in the forest population and in its surrounding environment, is pivotal to securing a timely response to emerging threats posed to forest and plant health.
Project FOCUS (Forest Operational monitoring using Copernicus and UAV hyperSpectral data), funded by the European Commission through the Horizon 2020 programme, aims at introducing new methodologies to innovatively address these challenges. To this end, new field-validated algorithms, based on a novel combination of satellite-derived multispectral data with UAV-derived hyperspectral data, are utilized to help detect changes in the forest population at an earlier stage than what is traditionally possible. Detecting these changes earlier allows for an improved mitigation strategy, and ultimately leads to a more sustainable forest resource and less value lost per year. The novel algorithms are integrated within the forest monitoring service Silvisense (www.silvisense.com), utilizing the latest within machine-learning methodologies to ensure higher quality results while maintaining a cost-efficient production.
A major threat to maritime pine (Pinus pinaster) trees is the infection by Pinewood Nematode (PWN) (Bursaphelenchus xylophilus), the biotic agent responsible for Pine Wilt Disease. Native to North America, B. xylophilus requires the simultaneous presence of a vector insect for propagation and suitable host trees (coniferous). The first detection of Pine Wilt Disease (PWD) in Europe happened in 1999, when it was identified in Portugal. The initial cases were followed by the fast spread of this disease nationwide, leading to a decrease of P. pinaster area. P. pinaster is a very important tree species for the Portuguese economy, through the use of its products and sub-products originated from pine wood. Other susceptible coniferous occur over wide areas in Europe and the potential ecologic and economic losses in the event of a continued expansion of the range of this biotic agent are significant. To contain PWD to Portugal and avoid the propagation of a major threat to European forests, special measures were adopted by the European Union and Portuguese governments. Since then, detections have been made outside of contiguous Portugal, including in the Madeira Islands and Spain, highlighting the serious challenges to containment and eradication efforts.
In PWD external symptoms include a decrease in resin exudation, browning and wilting of needles, and drying of the crown. In this disease, the decline process is very quick and also leads to changes in several parameters even before detection is visually possible.
Plants in stress conditions, produce less chlorophyll during the senescence process, this fact affects the light absorption, and causes the appearance of chloroses in plant tissues, and consequently changes in the reflectance captured from multi- and hyperspectral data. These facts contribute to the use of remote sensing technologies in support of timely detection efforts. Traditional surveying techniques are ineffective since it is not possible to identify infected trees in dense stands or remote areas from the ground. This limitation has greatly contributed to the unstoppable expansion of the biotic agent. Early detection through remote surveillance is pivotal to ensure European forests outside of the Iberian Peninsula remain PWD-free.
In this work, we present preliminary results of Very High Resolution and Sentinel-2 performance in the detection of the PWD, supported by laboratory and in situ spectroradiometry. An intensive ground survey campaign and multivariate laboratory analysis are also discussed. Healthy and infected trees were selected and monitored across a wilting season in a set of test sites in central Portugal. Several of the previously healthy trees were naturally infected in the course of the season, offering a unique perspective into the spectral and physiological evolution of the disease in the natural environment Monthly sample collection allowed the laboratorial analysis of multiple parameters including the quantification of pigments, phenolic compounds, and nitrates, as well as water content measurements. Correlations between remote, in situ, and laboratorial results were found and are now discussed for the first time.
Preliminary results strongly suggest high resolution satellite-based detection of wilting adult maritime pines is feasible in the context of European forests.

Intro/background:
The global forest economy is subject to a number of threats to its production value. Forest disease outbreaks become more frequent due to globalisation and climate change - they are difficult to contain and mitigate. According to FAO, in the period 1980-2002 more than 52 mill. hectares of forest in 37 countries were damaged by pests. The Pine Wilt Nematode is estimated, due to predictions in changing climate, to have the potential to spread to 34% of Europe by 2030. Forest fires are escalating in severity, and costs due to climate change are increasing. Europe lost in 2017 three times as large an area to forest fires than during the period 2008-2016 in total. The forest fires cost Portugal alone more than 200 mill. EUR and killed at least 66 people in 2017. These forest threats require accurate, precise and frequent information for monitoring tree stand status and planning any relevant mitigation actions. Sentinel-2 becoming operational in 2015, and the impressive results of deep learning techniques, served as an ideal starting point for the automated forest monitoring service Silvisense. Silvisense provides forest owners with a digital service to automatically monitor and report the state of forest inventory.

Challenge:
Early symptoms of forest disease are only reflected by small changes in the spectral reflectance, which are difficult to detect from Sentinel-2 imagery alone. Such minor variations in the spectral reflectance could be detected from airborne hyperspectral data, increasing the capacity to reveal disease outbreaks at an earlier stage.

Goal:
This forms the basis of the European Commission H2020 project FOCUS (Forest Operational monitoring using Copernicus and UAV hyperSpectral data), which aims to improve and extend the Silvisense forest monitoring services by combining multispectral data from Sentinel-2 with hyperspectral data from remotely piloted airborne platforms. The added value of the hyperspectral data particularly lies in the possibility to detect forest diseases at an earlier stage. The focus of this paper will be on the adaption of the Silvisense processing chain for early onset forest disease detection.

Methods:
An initial case study is conducted in Portugal, where outbreaks of the Pine Wilt Disease (PWD) threaten 70% of the forest population. A number of test sites have been selected at different geographical settings. At each site, tree samples are regularly taken for analyses in the lab. Their physical and biochemical variables are measured to find the effects of wilting within the pine needles. Spectral measurements enable to relate wilting to changes in the reflected spectra of the needles. In multiple airborne campaigns all test sites are visited to gather multi and hyperspectral data throughout the season. Aerial campaigns are synchronized with Sentinel-2 overpasses and ground sampling campaigns, such that the data can be related to each other. The resulting datasets serve to create accurate and high-resolution ground truth data for the parameters linked to PWD. Most importantly, this data can be used as training data for deep learning algorithms for Sentinel-2 image classification. The classified images are further improved by means of Markov smoothing over the temporal dimension. Since Sentinel-2 has a revisit time of 5 days, there is a lot to gain by exploiting temporal information. The actual mapping of disease outbreaks is based on change detection between classified images from different times. In order to efficiently process these large data volumes a processing framework is designed. One of the main advantages is, that it allows to reuse previously computed results from different components. This processing framework is part of the Silvisense forest monitoring service, in which the resulting products of the FOCUS project are implemented.

Preliminary results:
From the ground sampling and airborne campaigns already conducted within the FOCUS project, significant differences have been found between the reflected spectra of healthy and wilting pine needles. Furthermore, these differences can clearly be identified in the collected hyperspectral airborne data. Based on this, it is expected that high quality ground truth data can be created for training the Sentinel-2 image classifier. Ultimately this should make it possible to adapt the Silvisense processing chain for the early detection of PWD. This case study furthermore forms a proof of concept for the detection of similar symptoms from outbreaks of other forest diseases.

There is an urgent need for more accurate and timely information on changes and activities in tropical forest to empower effective law enforcement actions against illegal deforestation. With the Copernicus Sentinel-1 radar and Sentinel-2 optical satellites dense and regular observations with a spatial resolution of 10 m are provided globally for the first time. This provides an unprecedented potential to track forest changes and activities at high spatial resolution and in near real-time. Here, we present results from our recent work on radar-based (Sentinel-1) and multi-sensor (Sentinel-1 + Sentinel-2) methods for near real-time forest change alerting and characterization (Reiche et al., 2018a/b) undertaken across a number of tropical forest sites, including Indonesia, Bolivia and Peru. We will present an outlook on upcoming activities and future research directions.
Our results for Indonesia showed that dense Sentinel-1 data allow for confident and timely forest cover loss detection in natural and plantation forest with user’s and producer’s accuracy above 95%. Forest cover loss was detected and confirmed within three weeks in natural forest and within two weeks in plantation forest, a difference that can primarily be related to different change processes and dynamics in natural and plantation forest.
We show how environmental effects (e.g. rain) and seasonality affect the Sentinel-1 C-band radar signal and provide methods and concepts on how we can deal with it. We combined Sentinel satellite observations with frequently updated spatial data (e.g. logging roads and fires) to decrease the commission error and improve the timeliness of forest change alerting.
We combined near real-time Sentinel-1 and-2 based forest cover change information with active fire hotspot alerts from e.g. VIIRS and MODIS. Knowing the temporal accuracy and related uncertainties of the Sentinel based alerts, allowed us to characterize the temporal relationship between fires and tropical forest cover loss at high temporal detail and medium spatial scale for the first time. For Riau, Indonesia, we found clear spatial patterns of fires predating, coinciding or postdating forest cover loss patches.
We provide a more robust estimation of the temporal detection accuracy by improving the precision of the reference data derived from the multi-sensor satellite time series. We quantified how the near real-time forest cover change alerting is associated with a trade-off between the confidence in detection and the temporal detection accuracy. We showed that the trade-off affects the choice on how to use the near-real time data for different applications such as fast alerting with high temporal accuracy but lower confidence versus accurate detection at lower temporal detail.

Reiche, J., et al. (2018b) Characterizing Tropical Forest Cover Loss Using Dense Sentinel-1 Data and Active Fire Alerts. Remote Sensing, 10, 5, 777, doi:10.3390/RS10050777.
Reiche, J., et al. (2018a): Improving near-real time deforestation monitoring in tropical dry forests by combining dense Sentinel-1 time series with Landsat and ALOS-2 PALSAR-2. Remote Sensing of Environment, 204, 147-161. http://dx.doi.org/10.1016/j.rse.2017.10.034

Wildfires damage property and infrastructure, displace people, burn vegetation, and cost millions of dollars to fight. Timely and accurate information on fire perimeters, active wildfire front, and fire progression is critical for effective emergency management during wildfires. For active wildfire monitoring, the Moderate Resolution Imaging Spectroradiometer (MODIS) Active Fire maps are often used for contextual awareness while Landsat data are often deployed for post-wildfire boundary determination and burn severity mapping. Synthetic Aperture Radar (SAR) systems, capable to penetrate clouds and haze, have been used for fire scar mapping in previous studies. However, the use of SAR data for monitoring wildfire progression has been very limited due to the availability of SAR dense time series. With the launches of the ESA Sentinel-1, Synthetic Aperture Radar (SAR) with global coverage and frequent revisits become freely available. Sentinel-1 SAR data present excellent opportunity to develop a framework for SAR-based near real-time wildfire monitoring.

The overall objective of this research is to investigate Sentinel-1 SAR dense time series for near-real time wildfire progression mapping. Three long-duration wildfires are selected for this research: Elephant Hill Fire in British Columbia, Mendocino Complex Fire in Northern California and Thomas Fire in Southern California. An automatic method was developed in Google Earth Engine to analyze the backscatter trends of Sentinel-1 SAR time series and detect wildfire automatically. Continuous change detection was performed to monitor wildfire progression using a log ratio operator and a deep learning based on convolutional neural networks (CNN). The results show that, using Sentinel-1 time series data, wildfires could be detected and temporal progressions could be captured (Fig. 1). VH performed better and more stable in distinguishing burnt areas than VV, but VH and VV together achieved much better results than VV or VH alone . The results demonstrated that multitmeporal spaceborne SAR data is very promising for near real-time or real-time wildfire monitoring when SAR data is available on a daily or hourly basis with the planned launches of ICEYE and Capella Space Satellite Constellations.

Abstract:
Storm events with strong wind gusts can cause severe windthrows to large forest areas. A rapid detection of the spatial distribution of the windthrown areas is crucial for forest managers to help them direct their limited resources. In contrast to optical remote sensing data, Synthetic Aperture Radar (SAR) data is acquired largely independent of daylight or weather conditions. Thus, SAR sensors can produce temporally consistent and reliable data with a high revisit rate. In the present study, a straightforward approach was developed that uses Sentinel-1 (S-1) C-band VV and VH polarisation data for a rapid windthrow detection in mixed temperate forests for two study areas in Switzerland and northern Germany. First, several S-1 acquisitions of approximately 10 before and 30 days after the storm event were radiometrically terrain corrected [1, 2]. Second, based on these S-1 acquisitions, a SAR composite image of before and after the storm was generated. Then, after analysing the differences in backscatter between before and after the storm for windthrown and intact forest areas, a change detection method was developed to suggest potential locations of windthrown areas of a minimum extent of 0.5 ha – as it is required by the forest management (Figure 1). The detection is based on two user-defined parameters.
Overall, the results confirmed that the locations of most areal windthrows were successfully detected, however always slightly underestimated in area. While results from the independent study area in Germany indicated that the method is very promising for detecting areal windthrow with a producer’s accuracy of 0.88, its performance was less satisfactory at detecting single scattered windthrown trees. Moreover, the rate of false positives was low, with a user’s accuracy of 0.85 for (combined) areal and scattered windthrown areas.
With our study we show that C-band backscatter data have great potential to rapidly detect the locations of windthrow in mixed temperate forests within a short time (approx. 2 weeks) after a storm event. Furthermore, the two adjustable parameters allow a flexible application of the method tailored to the user’s needs [3].
Moreover, with the launch of the RADARSAT Constellation Mission (RCM) planned in early 2019, an increased number of SAR C-band acquisitions will be potentially available soon. Combined with S-1 data, the time period could be substantially shorter to receive the number of required acquisitions for a windthrow detection of similar quality. This would make the approach using reliable SAR data worthy of consideration, even if its performance is worse than methods using more expensive ALS or optical data with longer latencies due to meteorological conditions. Forest managers would greatly benefit from an immediate first overview map of the windthrow locations, as it allows them to begin their planning of the clearing and reforestation of the windthrown areas as soon as possible.

Figure:
Example of the functionality of the windthrow detection method. (a) True color Planet image of the region with the borders of the digitised windthrown areas (yellow lines) and (b) windthrow index (WI) of the region in dB. (c) The result after the first branch in the decision tree with all potential windthrown areas in the forest. (d) The final result after application of the minimum size threshold. Only larger objects remain, leading to a more realistic representation of the windthrow in the forest.

References:

[1] Small, D., “Flattening Gamma: Radiometric Terrain Correction for SAR Imagery,” IEEE Trans. Geosci. Remote Sens., 49(8), 3081–3093, Aug. 2011.
[2] Small, D., “SAR backscatter multitemporal compositing via local resolution weighting,” in IEEE International Geoscience and Remote Sensing Symposium, Munich, Germany, 2012, 4521–4524.
[3] Rueetschi, M., Small, D., Waser, L.T., Rapid Detection of Windthrows Using Sentinel-1 C-band SAR Data. in review for Remote Sensing