Forestry Methods 1

Detecting deforestation in near real-time (NRT) is essential to effectively manage and protect forest resources in the tropics. Current remote sensing based NRT monitoring systems rely on MODIS time series imagery in order to provide fortnightly information on newly deforested areas at 500m resolution. Due to the low spatial resolution, however, small scale changes are missed, which precludes the rapid response of many human-induced deforestation activities that tend to be small scale. Missing data due to persistent cloud cover limtes optical-based NRT monitoring at scale medium resolution scale.

We present a Bayesian approach to combine SAR and optical time series for near real-time deforestation detection at medium resolution scale (Reiche et al. 2015). Once a new image of either of the input time series is available, the conditional probability of deforestation is computed using Bayesian updating, and deforestation events are indicated. Future observations are used to update the conditional probability of deforestation and, thus, to confirm or reject an indicated deforestation event.

Results are presented using (i) Landsat and ALOS PALSAR time series of an evergreen forest plantation in Fiji, where we emulated a near real-time scenario. Three-monthly reference data (plantation operations) covering the entire study areas was used to validate and assess spatial and temporal accuracy. We tested the impact of persistent cloud cover by increasing the per-pixel missing data percentage of the NDVI time series stepwise from ~50% (~6 observations/year) up to 95% (~0.5 observations/year) while combining with a consistent PALSAR time series of ~2 observations/year. Spatial and temporal accuracies for the fused Landsat-PALSAR case (overall accuracy = 87.4%; mean time lag of detected deforestation = 1.3 months) were consistently higher than those of the Landsat- and PALSAR-only cases. The improvement maintained even for increasing missing data in the Landsat time series.

Preliminary results of using dense Sentinel-1 C-band time series in addition to Landsat and ALOS PALSAR-1 and -2 time series will be presented for a forest sites in Ethiopia and Bolivia. These results in particular highlight the capability of dense Sentinel-1 C-band time series for twice-weekly deforestation monitoring.

Reiche, J., de Bruin, S., Hoekman, D. H., Verbesselt, J. & Herold, M. A Bayesian Approach to Combine Landsat and ALOS PALSAR Time Series for Near Real-Time Deforestation Detection. Remote Sens. 7, 4973–4996 (2015).

Recent developments in remote sensing technology have improved our abilities of forest monitoring remarkably. Accurate 3D scans of forest canopies can be obtained from airborne light detection and ranging (Lidar) and increasingly from synthetic aperture Radar (SAR) satellites (e.g. TanDEM-X, Sentinel-1, BIOMASS). Thus, information on forest heights can be gathered and mapped at large scales and different spatial resolutions. However, the main variable of interest, e.g. in the context of the UN's initiative for reduction of emissions from deforestation and forest degradation (REDD+), is biomass not height. Further variables which are crucial in understanding the role of forests in the global carbon cycle are productivity, carbon turnover, basal area, stem size distribution, forest age and disturbance regimes. By analyzing inventory data regression models between remote sensing metrics and the intrinsic target variables can be established. For example, canopy height usually forms a strong relationship with above ground biomass. However, the inventory-based calibration requires extensive field campaigns and reaches its limits with variables which are difficult to measure at plot scale, like carbon fluxes.

Here, we propose a new approach where we complement real world ground-truth data with virtual forest inventory data generated by forest simulations. Dynamic forest models help us to understand the links between ecological processes and vegetation structure. Hence, joining the fields of remote sensing and forest modelling will lead to mutual benefits, by increasing structural realism of forest models and facilitating the development and calibration of remote sensing approaches.

For this purpose, we developed a Lidar simulator for the individual-based forest model FORMIND. The target of the first application study was finding good Lidar-based predictors of above ground biomass for a tropical rainforest. Various Lidar metrics have served for forest biomass estimation in the past, with the calibration of the Lidar-to-biomass-relationships traditionally relying on ground-truth data collected manually in field inventory plots. We, instead, used output of FORMIND simulations as virtual inventory data and simulated Lidar measurements resulting in virtual remote sensing data. The tropical rainforest on Barro Colorado Island (BCI), Panama, for which extensive validation data (inventories and Lidar) is available, served as a study site. With the forest model it is fairly easy to generate data that covers the entire range of possible biomass stocks and forest successional states by including typical disturbances in the simulation. This is a big advantage over field inventories which are often restricted. Good biomass predictions were obtained with several different Lidar-metrics, including some that had not been applied on tropical forests before.

Future applications of this modelling approach could go further from static biomass stock to dynamic carbon flux estimation via remote sensing. Furthermore, the methodology can be transferred to SAR observations. This approach allows exploration of the potential of current and future satellite missions (e.g. Tandem-L) for estimating intrinsic forest attributes.

The Brazilian Amazon is one of the world’s major deforestation hotspots, although deforestation rates reported by the Instituto Nacional de Pesquisas Espaciais show a decreasing trend from its highest value in the mid-1990s (~30,000 km² yr^-1) to a record low of ~4,500 km² yr^-1 in 2012. Abandoned croplands and pastures in tropical regions give rise to rapid establishment of secondary forests, with consequent carbon accumulation and potential restoration of biodiversity. However, the age, type and composition of these forests established on abandoned lands are a consequence of several factors, such as land use history, soil fertility, and distance to mature (primary) forests.

The land use history prior to land abandonment can be retrieved by analysing time series of land cover maps derived from classification of high resolution optical data. However, this approach is often hindered by cloud cover, leading to poor coverage by optical data. All-weather SAR data availability is increasing with consequent benefits for land use/land cover change monitoring over tropical regions.

The objective of this study was to assess the capability of optical and L-band SAR data to discriminate stages of tropical forest regeneration across the Brazilian Amazon. Available 3-class (mature forest, non-forest, secondary forest) land cover maps over three selected sites of the Brazilian Amazon (Manaus: 1973-2011; Santarém: 1984-2010; Machadinho d’Oeste: 1984-2011) were used to generate maps of age of secondary forests (ASF) at any given year in those periods. The ability of ALOS PALSAR backscatter intensity and Landsat TM surface reflectance to discriminate stages of tropical forest regeneration was tested using data acquired over those sites between 2007 and 2010.

Models of the type I = a₀ + a₁ ln(ASF) (where I = HH or HV backscatter intensity, surface reflectance or vegetation indices) were fitted to the data from Manaus; preliminary results indicate that only the models using HH backscatter intensity, TM band 3 (red) or Normalised Difference Vegetation Index (NDVI) were not significant for an α=0.05. Model goodness of fit (r²) was low (0.01-0.05) but marginally better when using TM band 4, Soil Adjusted Vegetation Index (SAVI), Modified SAVI (MSAVI), Enhanced Vegetation Index (EVI) and HV backscatter intensity. However considerable inter-annual variability is noticeable. The analysis of variance of the model relating I  to ASF classes at Manaus showed that only surface reflectance bands or vegetation indices were capable of discriminating ASF classes for an α=0.05. Again, substantial variability was observed across the 2007-2010 period.

Most of the forests regenerating at Manaus are older than 15 years and, therefore, displaying already a vegetation structure more similar to mature (primary) forests. Therefore, it’s not unexpected that surface reflectance data was better at discriminating ASF and ASF classes. It is anticipated that the analysis of younger forest regeneration, as found in Santarém and Machadinho d’Oeste, will provide better understanding of the capability of L-band radar data to discriminate tropical forest regeneration in the Brazilian Amazon.

Given the importance of forest ecosystems, the availability of reliable and spatially explicit information about the site-specific climate sensitivity of tree species is essential for implementing suitable adaptation strategies in forest management. Hyperspectral airborne data was used to detect drought sensitive areas within a deciduous forest in South West Germany, and results were compared with leaf measurements and soil moisture regime maps.

In a first campaign in July 2014, airborne hyperspectral data from a HySpex sensor was used to assess the response of deciduous species (dominated by European beech and Sessile and Pedunculate oak) to water stress during a summery dry spell. Spectral indices were used to evaluate whether the spectral response of trees under different ecological growing conditions is already distinguishable at an early stage of drought stress.

After masking canopy gaps, shaded crown areas and non-deciduous species, potentially indicative spectral indices (photochemical reflectance index (PRI), moisture stress index (MSI), normalized difference water index (NDWI) and chlorophyll index (CI)) were analysed with respect to available maps of site-specific soil moisture regimes. PRI provided important indication of current site-specific photosynthetic stress on leaf level in relation to limitations in soil water availability. The CI, MSI and NDWI revealed statistically significant differences in total chlorophyll and water concentration on canopy level. However, after reducing the canopy effects by normalizing these indices with the structure-sensitive simple ratio (SR) vegetation index, it was not yet possible to identify site-specific concentration differences on leaf level. The selected indicators were also tested with simulated EnMAP and Sentinel-2 data (derived from the original airborne data set). While PRI proved to be useful also on the spatial resolution of EnMAP (GSD = 30 m), PRI could not be reproduced with Sentinel-2, owing to the lack of adequate spectral bands. The remaining indicators (MSI, CI, SR) were successfully produced also with Sentinel-2 data at superior spatial resolution (GSD = 10 m). This campaign confirmed the importance of using earth observation systems for supplementing traditional ecological site classification maps, particularly during dry spells and heat waves when ecological gradients are increasingly reflected in the spectral response at tree crown level. It also underlined the importance of using Sentinel-2 and EnMAP in synergy, as soon as both systems will be available.

A second campaign was conducted in August 2015 when an airborne data set was acquired and ground measurements were taken. Tree climbers sampled 240 leaves from 13 trees under different growing conditions. Only leaves in the upper crown were collected, because these contribute most to the spectral signal in the sensor. Reference measurements were used to compare the spectral indices with laboratory measurements of chlorophyll, water and carotenoid content as well as C/N ratio. Additionally, spectral measurements of the sampled leaves with a field spectroradiometer were used to model canopy reflectance and use chlorophyll, water content and PRI values to create maps of current water stress as well as a map highlighting stands with high risk of drought stress.

Results from previous studies showed that reliable maps of leaf water and chlorophyll content could be produced on leaf level. First results on canopy level, using hyperspectral indices, and the availability of reference data enable to transfer these methods from leaf to canopy scale and can contribute to filling the gap between field imaging spectroscopy and the assessment of ecophysiological conditions of large forest communities.

The dryness stress situation in late summer 2015 gives the possibility to test if simulated EnMAP data still exceeds simulated Sentinel-2 data, or if both sensors can be used to detect drought stress in an advanced stage.

The Copernicus program of ESA and European commission (6 Sentinels Missions, among them Sentinel-1 with Synthetic Aperture Radar sensor and Sentinel-2 with 13-band 10 to 60 meter resolution optical sensors), offers a new opportunity to Earth Observation with high temporal acquisition capability (~12 days repetitiveness and ~5 days in some geographic areas of the world) with high spatial resolution.

Due to these high temporal and spatial resolutions, it opens new challenges in several fields such as image processing, new algorithms for Time Series and big data analysis.  In addition, these missions will be able to analyze several topics of earth temporal evolution such as crop vegetation, water bodies, Land use and Land Cover (LULC), sea and ice information, etc. This is particularly useful for end users and policy makers to detect early signs of damages, vegetation illness, flooding areas, etc.

From the state of the art, one can find algorithms and methods that use a bi-date comparison for change detection [1-3] or time series analysis. Actually, these methods are essentially used for target detection or for abrupt change detection that requires 2 observations only.

A Hölder means-based change detection technique has been proposed in [2,3] for high resolution radar images. This so-called MIMOSA technique has been mainly dedicated to man-made change detection in urban areas and CARABAS – II project by using a couple of SAR images. Hölder means definition for Time Series H_p[T] : N images * S Bandes * t dates = {x_i}_{i \in {1, 2, …, N},}

H_p[T] = (1/n (Somme _{i=1 to n} (X_p))^^(1/p), n \in {1, 2, …, N}

H_p (T) is continuous and monotonic increasing in p for − ∞ < p  < ∞

An extension to multitemporal change detection technique has been investigated but its application to land use and cover changes still has to be validated. The study of change detection in Time Series Analysis is possible using Holder means in multi spectral and SAR images from new missions like Sentinel-1, Sentinel-2, RadarSat Constellation (2017), NISAR – ISRO Mission (2020). This study is often done by computing and analyzing the Hölder factor (p) and provides multitemporal change detection for several applications.

The purpose of this paper is to take into consideration the entire time series for change detection or temporal evolution analysis (without requiring bi-date comparisons), for water bodies (water level), forest and floods areas in several sites in America, Africa and China. These test sites, bellowing to SPOT Take 5 experiment 2015 organized by CNES, CESBIO and ESA, have been chosen because of there thematic interest and also because these site present a high rate of successful acquisition of free clouds images, (percentage of clouds < 10 %). , with a ranging of 9 to 21 acquisitions almost free of clouds between  April to September 2015. These Sentinel  data (same resolution 10m as Sentinel2, 5 days revisit, an almost a SWIR band)  have been delivered by CNES as  Level 2A products (Ortho & TOC reflectance) .

At this date, we have tested some sites that present similar characteristics in landscape structure like river or lakes, forest (Novo Progresso in Brasil, Waku Kungo in Angola) and sometime holding high temporal flooding areas (Poyang lake in China, La Victoria - Magdalena River/Wetlands in Colombia).

Using the main spectral properties of each band (SWIR for water characterization) or groups of bands (SWIR and NIR for water bodies and floods areas), we can make the most of Hölder means methods in Long Time Data Series for change detection in water bodies (water Level) and floods areas.

On the other hand, we can use all bands and all dates in Long Time Data Series for Temporal LULC classification.

Our algorithm showed very robust performance on several test sites. In Poyang Lake (China), La Victoria (Colombia) we have extracted temporal change mainly for water bodies (water extent) and floods areas. For Novo Progresso (Brasil) and Waku Kungo (Angola) we have performed the change detection in water bodies and forest, using time series of SPOT (Take 5) project, Sentinel-1 and Sentinel-2 images.

At this point, we can foresee the option of fusion of Holder means analysis derived from optical-radar data, for better results in temporal change detection applications. This option allows to improve performance when the optical imagery has a high clouds cover. Holder method in Long Time Data Series also has further applications for forest fire, LULC, etc.

-----

[1] Niculescu, S., Lardeux, C., Hanganu, J., Mercier, G., and David, L. (2013). Change detection in floodable areas of the Danube delta using radar images. Natural Hazards, 1-18.

 [2] Quin, G. Etude des séries temporelles en imagerie satellitaire SAR pour la détection automatique de changements, Thèse de doctorat, Telecom ParisTech, 27 janvier 2014.

[3] Quin, G. Pinel-Puyssegur, B. and Nicolas, J-M. Comparison of Harmonic, Geometric and Arithmetic means for change detection in SAR time series, in Synthetic Aperture Radar, 2012. EUSAR. 9^th European Conference on. VDE, 2012, pp. 255–258.