Search

Identifying and characterizing vascular plants in time and space is required in various disciplines, e.g. in forestry, conservation and agriculture. Remote sensing emerged as a key technology revealing both spatial and temporal vegetation patterns. Harnessing the ever growing streams of remote sensing data for the increasing demands on vegetation assessments and monitoring requires efficient, accurate and flexible methods for data analysis. In this respect, the use of deep learning methods is trend-setting, enabling high predictive accuracy, while learning the relevant data features independently in an end-to-end fashion. Very recently, a series of studies have demonstrated that the deep learning method of Convolutional Neural Networks (CNN) is very effective to represent spatial patterns enabling to extract a wide array of vegetation properties from remote sensing imagery. This review introduces the principles of CNN and distils why they are particularly suitable for vegetation remote sensing. The main part synthesizes current trends and developments, including considerations about spectral resolution, spatial grain, different sensors types, modes of reference data generation, sources of existing reference data, as well as CNN approaches and architectures. The literature review showed that CNN can be applied to various problems, including the detection of individual plants or the pixel-wise segmentation of vegetation classes, while numerous studies have evinced that CNN outperform shallow machine learning methods. Several studies suggest that the ability of CNN to exploit spatial patterns particularly facilitates the value of very high spatial resolution data. The modularity in the common deep learning frameworks allows a high flexibility for the adaptation of architectures, whereby especially multi-modal or multi-temporal applications can benefit. An increasing availability of techniques for visualizing features learned by CNNs will not only contribute to interpret but to learn from such models and improve our understanding of remotely sensed signals of vegetation. Although CNN has not been around for long, it seems obvious that they will usher in a new era of vegetation remote sensing.

Xylella fastidiosa (Xf) is a harmful plant pathogenic bacterium, able to infect over 500 plant species worldwide. Successful eradication and containment strategies for harmful pathogens require large-scale monitoring techniques for the detection of infected hosts, even when they do not display visual symptoms. Although a previous study using airborne hyperspectral and thermal imagery has shown promising results for the early detection of Xf-infected olive (Olea europaea) trees, further work is needed when adopting these techniques for large scale monitoring using multispectral cameras on board airborne platforms and satellites. We used hyperspectral and thermal imagery collected during a two-year airborne campaign in a Xf-infected area in southern Italy to assess the performance of spectrally constrained machine-learning algorithms for this task. The algorithms were used to assess multispectral bandsets, selected from the original hyperspectral imagery, that were compatible with large-scale monitoring from unmanned platforms and manned aircraft. In addition, the contribution of solar–induced chlorophyll fluorescence (SIF) and the temperature-based Crop Water Stress Index (CWSI) retrieved from hyperspectral and thermal imaging, respectively, were evaluated to quantify their relative importance in the algorithms used to detect Xf infection. The detection performance using support vector machine algorithms decreased from ∼80% (kappa, κ = 0.42) when using the original full hyperspectral dataset including SIF and CWSI to ∼74% (κ = 0.36) when the optimal set of six spectral bands most sensitive to Xf infection were used in addition to the CWSI thermal indicator. When neither SIF nor CWSI were used, the detection yielded less than 70% accuracy (decreasing κ to very low performance, 0.29), revealing that tree temperature was more important than chlorophyll fluorescence for the Xf detection. This work demonstrates that large-scale Xf monitoring can be supported using airborne platforms carrying multispectral and thermal cameras with a limited number of spectral bands (e.g., six to 12 bands with 10 nm bandwidths) as long as they are carefully selected by their sensitivity to the Xf symptoms. More precisely, the blue (bands between 400 and 450 nm to derive the NPQI index) and thermal (to derive CWSI from tree temperature) were the most critical spectral regions for their sensitivity to Xf symptoms in olive., Data collection was partially supported by the European Union’s Horizon 2020 research and innovation program through grants to the POnTE (Pest Organisms threatening Europe; grant 635646 from European Union’s Horizon 2020 Framework Research Programme) and XF-ACTORS (Xylella fastidiosa Active Containment Through a Multidisciplinary-Oriented Research Strategy; grant 727987 from European Union’s Horizon 2020 Framework Research Programme) projects. A. Hornero was supported by a research fellowship DTC GEO 29 “Detection of global photosynthesis and forest health from space” from the Science Doctoral Training Centre (Swansea University, UK).

Invasive plant species can pose major threats to biodiversity, ecosystem functioning and services. Satellite based remote sensing has evolved as an important technology to spatially map the occurrence of invasive species in space and time. With the new era of the Sentinel missions, Synthetic Aperture Radar (SAR) and multispectral data are now freely available and repeatedly acquired on a high spatial and temporal resolution for the entire globe. However, the high potential of such sensors for automatic mapping procedures cannot be fully harnessed without sufficient and appropriate reference data for model calibration. Reference data are commonly acquired in field surveys, which however, are often relatively expensive and affected by sampling and observer bias. Moreover, a direct transferability to the remote sensing perspective and scale is difficult. Accordingly, we firstly assess the potential of Unmanned Aerial Vehicles (UAV) for semi-automatic reference data acquisition on species cover of three woody invasive species Pinus radiata, Ulex europaeus and Acacia dealbata occurring in Chile. Secondly, we test the upscaling of the estimated species cover to the spatial scale of Sentinel-1 and Sentinel-2. The proposed workflow includes the visual sampling of respective canopies in UAV orthomosaics and the subsequent spatial extrapolations using MaxEnt with spectral (RGB, Hyperspectral), textural (2D) and canopy structural (3D) predictors derived from UAV-based photogrammetry. These UAV-based maps are then used to train random forest models with multitemporal Sentinel-1 and Sentinel-2 data to map the invasive species cover on large spatial scales. Our results show that the semi-automatic UAV-based mapping of the three invasive species results in accurate predictions. Depending on the predictor combination, the correlation was 0.70, 0.77 and 0.90 for Pinus radiatia, Ulex europaeus, Acacia dealbata, respectively. Among the three species, we observed clear differences in the model performance between the tested photogrammetric predictors and their combinations (spectral, 2D texture or 3D structure). For scaling up the UAV-based estimates to the satellite-scale, the Sentinel-2 data (multispectral) were more important than Sentinel-1 data (SAR). An independent validation revealed that the R2 of the upscaling accounted for 0.78 or higher for all species and RMSE lower than 12%. Our results hence demonstrate that UAV-based reference data acquisitions are a promising alternative to traditional field surveys if the target species are directly identifiable in the UAV data.

Deep learning and particularly Convolutional Neural Networks (CNN) in concert with remote sensing are becoming standard analytical tools in the geosciences. A series of studies has presented the seemingly outstanding performance of CNN for predictive modelling. However, the predictive performance of such models is commonly estimated using random cross-validation, which does not account for spatial autocorrelation between training and validation data. Independent of the analytical method, such spatial dependence will inevitably inflate the estimated model performance. This problem is ignored in most CNN-related studies and suggests a flaw in their validation procedure. Here, we demonstrate how neglecting spatial autocorrelation during cross-validation leads to an optimistic model performance assessment, using the example of a tree species segmentation problem in multiple, spatially distributed drone image acquisitions. We evaluated CNN-based predictions with test data sampled from 1) randomly sampled hold-outs and 2) spatially blocked hold-outs. Assuming that a block cross-validation provides a realistic model performance, a validation with randomly sampled holdouts overestimated the model performance by up to 28%. Smaller training sample size increased this optimism. Spatial autocorrelation among observations was significantly higher within than between different remote sensing acquisitions. Thus, model performance should be tested with spatial cross-validation strategies and multiple independent remote sensing acquisitions. Otherwise, the estimated performance of any geospatial deep learning method is likely to be overestimated.

Leaf mass per area (LMA), leaf dry matter content (LDMC) and leaf water content/ equivalent water thickness (EWT) are commonly used functional plant traits in ecology. Whereas spectroscopy has recently proven to be a powerful tool to collect such functional trait information across large scales, it remains unclear whether these reflectance-based trait predictions are accurate enough to reliably model trait variation at the intraspecific level (i.e. across individuals of one species). We explored the potential of hyperspectral leaf reflectance-based methods to predict LMA, LDMC and EWT at the intraspecific level for two herbs (Hieracium umbellatum and Jacobaea vulgaris) and two shrubs (Rosa rugosa and Rubus caesius), based on 2400 leaf samples. More specifically we tested i) inversion of the PROSPECT-D radiative transfer model, ii) a generic PLSR approach using the multibiome LMA PLSR model and iii) a data-specific PLSR approach at the species level. For the latter approach we furthermore assessed both model transferability across species and the trade-off between sample size and model accuracy. Although the PROSPECT-D model inversion and the multibiome LMA PLSR model were relatively accurate for intraspecific LMA predictions of shrubs (R2 > 71 and 76%, respectively, however NRMSE = 33–47%), their performance was lower for herbs (R2 70%, NRMSE

Grasslands are one of the ecosystems that have been strongly affected by anthropogenic impacts. The state-of-the-art in monitoring changes in grassland species composition is to conduct repeated plot-based vegetation surveys that assess the occurrence and cover of plants. These plot-based surveys are typically limited to comparably small areas and the quality of the cover estimates depends strongly on the experience and performance of the surveyors. Here, we investigate the possibility of a semi-automated, image-based method for cover estimates, by analyzing the applicability of very high spatial resolution hyperspectral data to classify grassland species at the level of individuals. This individual-oriented approach is seen as an alternative to community-oriented remote sensing depicting canopy reflectance as the total of mixed species reflectance. An AISA + imaging spectrometer mounted on a scaffold was used to scan 1 m 2 grassland plots and assess the impact of four sources of variation on the predicted species cover: (1) the spatial resolution of the scans, (2) complexity, i.e. species number and structural diversity, (3) the species cover and (4) the share of functional types (graminoids and forbs). Classifications were conducted using a support vector machine classification with a linear kernel, obtaining a median Kappa of ~ 0.8. Species cover estimations reached median r 2 and root mean square errors (RMSE) of ~ 0.6 and ~ 6.2% respectively. We found that the spatial resolution and diversity level (mainly structural diversity) were the most important sources of variation affecting the performance of the proposed approach. A spatial resolution below 1 cm produced relatively good models for estimating species-specific coverages (r 2 = ~ 0.6; RMSE = ~ 7.5%) while predictions using pixel sizes over that threshold failed in this individual-oriented approach (r 2 = ~ 0.17; RMSE = ~ 20.7%). Areas with low inter-species overlap were better suited than areas with frequent inter-species overlap. We conclude that the application of very high resolution hyperspectral remote sensing in environments with low structural heterogeneity is suited for individual-oriented mapping of grassland plant species.

Spaceborne sensors allow for wide-scale assessments of forest ecosystems. Combining the products of multiple sensors is hypothesized to improve the estimation of forest biomass. We applied interferometric (Tandem-X) and photogrammetric (WorldView-2) based predictors, e.g. canopy height models, in combination with hyperspectral predictors (EO1-Hyperion) by using 4 different machine learning algorithms for biomass estimation in temperate forest stands near Karlsruhe, Germany. An iterative model selection procedure was used to identify the optimal combination of predictors. The most accurate model (Random Forest) reached a r 2 of 0.73 with a RMSE of 14.9% (29.4 t/ha). Further results revealed that the predictive accuracy depended highly on the statistical model and the area size of the field samples. We conclude that a fusion of canopy height and spectral information allows for accurate estimations of forest biomass from space.

Peatlands are key reservoirs of belowground carbon (C) and their monitoring is important to assess the rapid changes in the C cycle caused by climate change and direct anthropogenic impacts. Frequently, information of peatland area and vegetation type estimated by remote sensing has been used along with soil measurements and allometric functions to estimate belowground C stocks. Despite the accuracy of such approaches, there is still the need to find mappable proxies that enhance predictions with remote sensing data while reducing field and laboratory efforts. Therefore, we assessed the use of aboveground vegetation attributes as proxies to predict peatland belowground C stocks. First, the ecological relations between remotely detectable vegetation attributes (i.e. vegetation height, aboveground biomass, species richness and floristic composition of vascular plants) and belowground C stocks were obtained using structural equation modeling (SEM). SEM was formulated using expert knowledge and trained and validated using in-situ information. Second, the SEM latent vectors were spatially mapped using random forests regressions with UAV-based hyperspectral and structural information. Finally, this enabled us to map belowground C stocks using the SEM functions parameterized with the random forests derived maps. This SEM approach resulted in higher accuracies than a direct application of a purely data-driven random forests approach with UAV data, with improvements of r 2 from 0.39 to 0.54, normalized RMSE from 31.33% to 20.24% and bias from −0.73 to 0.05. Our case study showed that: (1) vegetation height, species richness and aboveground biomass are good proxies to map peatland belowground C stocks, as they can be estimated using remote sensing data and hold strong relationships with the belowground C gradient; and (2) SEM is facilitates to incorporate theoretical knowledge in empirical modeling approaches.

Photosynthesis is essential for life on earth as it, inter alia, influences the composition of the atmosphere and is the driving mechanism of primary production. Photosynthesis is particularly controlled by leaf pigments such as chlorophyll, carotenoids or anthocyanins. Incoming solar radiation is mainly captured by chlorophyll, whereas plant organs are also protected from excess radiation by carotenoids and anthocyanins. Current and upcoming optical earth observation sensors are sensitive to these radiative processes and thus feature a high potential for mapping the spatial and temporal variation of these photosynthetic pigments. In the context of remote sensing, leaf pigments are either quantified as leaf area-based content [μg/cm2] or as leaf mass-based concentration [g/g or %]. However, these two metrics are fundamentally different, and until now there has been neither an in-depth discussion nor a consensus on which metric to choose. This is notable considering the amount of studies that do not explicitly differentiate between pigment content and concentration. We therefore seek to outline the differences between both metrics and thus show that the remote sensing of leaf pigment concentration [%] is unsubstantial. This is due to the fact that, firstly, pigment concentration is likely to primarily reflect variation in leaf mass per area and not pigments itself. Second, the radiative transfer in plant leaves is especially determined by the absolute content of pigments in a leaf and not its relative concentration to other leaf constituents. And third, as a ratio, pigment concentration is an ambiguous metric, which further complicates the quantification of leaf pigments at the canopy scale. Given these issues related to the use of chlorophyll concentration, we thus conclude that remote sensing of leaf pigments should be primarily performed on an area basis [μg/cm2]., The project was funded by the German Aerospace Center (DLR) on behalf of the Federal Ministry of Economics and Technology (BMWi), FKZ50EE 1347.