Tree of heaven (Ailanthus altissima) belongs to the 100 worst European invasive species with great potential of spread due to climate change. The tree is particularly limited by late spring frost damaging the buds and young leaves. In Czech Republic and other Central and Northern European countries, Ailanthus is still mostly present in urban areas serving as heat islands. However, with increasing temperatures and enough propagules from already established populations, its invasion is expected to grow rapidly. Means of early detection as well as detailed knowledge on species potential spread across Europe are therefore urgently needed. In Czech Republic, the species is currently on the northern edge of its distribution range, being present only in the warmest parts of the country, and in urban areas. Examples of Ailanthus escape to the open landscape are still rather rare here, unlike in neighbouring countries to the South - Hungary or Southern Austria where it already massively increased in last decades. But we can expect similar scenarios.
In our study, we focused on possibilities of early detection using custom built fixed wing UAV and identified the best phenology phase for detection. We have generated an orthoimage and digital model of the canopy using Structure from Motion method in Agisoft Photoscan. Consequently, we have applied object-oriented image analysis and automatic classification of segments and achieved high accuracies using spring multispectral imagery (RGB+NIR). Our results from several years show, that the accuracy also depends on the frosts, that can lower the accuracy, which is problematic, if the frost appeared only in part of the area (e.g.: valley bottoms).
Consequently, we have produced a model of current and potential distribution of the species, using data on species distribution in Czech Republic and identified environmental and climate constrains shaping the current distribution, and potential spread with regard to expected future climate change.

Some invasive plant species cause major negative affects to biodiversity and ecosystem services. Remote sensing has been frequently used to map invasive species across space and time. With the raise of the new Copernicus fleet and particularly the launch of the Sentinel-1 and 2 satellites, unprecedented high quality data at high repetition rates and high spatial and temporal resolution are available for the entire globe. These data have the potential to notably improve the state-of-art in remote sensing of invasive species. However, to fully use the high potential of the Copernicus data, sufficient and appropriate reference data are required for model calibration. These data are typically acquired during field surveys which often limits the surveyed area. Furthermore, field data are measured with a perspective that widely differs from the bird-eye view of the satellites.
In this study, we examine whether data collected from Unmanned Aerial Vehicles (UAV) can serve to semi-automatically acquire reference data to be subsequently upscaled to the Sentinel level. In the study we focus on three woody invasive species (Pinus radiata, Ulex europaeus, Acacia dealbata) affecting large areas of central Chile. In our workflow, we first manually digitize individuals of the three invasive species from the UAV images as training data. Then, we apply MaxEnt trained with textural (2D), spectral (RGB, Hyperspectral), and structural (3D) predictors derived from UAV-based photogrammetry to create occurence maps from the UAV mosaics. The MaxEnt occurrence maps for the three species are then transformed into fractional cover estimates of the target species and used as reference in random forest models which are trained with multitemporal Sentinel-1 and -2 data to map the invasive species cover on large scales.
The semi-automatic UAV-based mapping of invasive species resulted in acceptable accuracies (Kappa of 0.70 0.77, 0.90 for Pinus radiatia, Ulex europaeus, Acacia dealbata, respectively). The up-scaling process proved to be very successful reaching R2 values of 0.95 and higher for all species. This indicates that UAV-based reference data can be up-scaled almost without any accuracy losses to the Sentinel-level. We believe that the suggested work-flow could be a highly relevant and transferrable tool for mapping woody invasive species (or other species of interest) across larger areas.

Mapping the occurrence patterns of invasive plant species and understanding their invasion dynamics is a crucial requirement for preventing further spread to so far unaffected regions.

An established approach to map invasive species across large areas is based on the combination of satellite or aerial remote sensing data with ground-truth reference data from fieldwork. Unmanned aerial vehicles (UAV) may represent an interesting and low-cost alternative data source to labor-intensive fieldwork. Despite the increasing use of UAVs in the field of remote sensing in the last years, operational methods are missing to combine UAV and satellite data. Here, we present a new methodological framework to estimate the fractional cover of the invasive shrub species Ulex europaeus L. (common gorse), based on ultra-high resolution UAV images and a medium resolution intra-annual time-series of Sentinel-2.

In our workflow we translate the challenging co-registration between the two datasets with strongly different spatial resolutions into an optimization problem where the UAV orthoimages were systematically shifted against the satellite images. Based on several local random forest (RF) models, an optimal fit between the varying Ulex europaeus cover fractions from the UAV orthoimages and the spectral information of Sentinel-2 was searched to find the spatial offsets between both datasets. Considering these spatial offsets, led to significant improvements for the global RF model that was applied to map the current distribution of Ulex europaeus for Chiloé Island (south-central Chile). Through this approach, it was possible to identify, that the average spatial offsets between UAV and satellite data was approximately 8 m on the x-axis and approx. 7 m on the y-axis. These errors are in accordance with the above-mentioned expected positional errors of standard on-board (navigation) GNSS modules. Taking these spatial offsets into account, the local Ulex europaeus prediction was improved by up to 10 %-RMSE

Based on a semi-exhaustive search technique, we furthermore found that the Sentinel-2 acquisitions from March and November (flowering time of Ulex europaeus) were particularly important in distinguishing Ulex europaeus from other plant species. Our mapping results could support local efforts in controlling Ulex europaeus. Furthermore, the proposed workflow should be transportable to other use-cases where individual target species that are visually detectable in UAV imagery are considered.

Tree height is an important ecological trait, and part of many natural resource data collections, being the most widely used indicator of a site’s suitability for a variety of stand management uses, ranging from wildlife habitat to timber production, and can also provide indications on forest health conditions and growth trends.
In computing above ground biomass (AGB) height information is crucial, with previous work demonstrating that the incorporation of height into allometric models in addition to the diameter variable significantly improves AGB estimates, especially in tropical forests. At present, most of the widely adopted models include the height variable.
Tree heights are difficult to measure in the field, especially in tropical forests where tall, closed canopies, and dense understory occur, limiting the sight of tree tops. When height is used as input to calculate AGB, inaccuracy in measures can lead to large errors in biomass estimates. Many techniques can be used to measure tree heights; here we focus on Airborne Lidar Scanning (ALS), and stereo-photogrammetry with imagery acquired by a RGB camera mounted on Unmanned Aerial Vehicle (UAV). These approaches increase the possibility of precise tree height measures while reducing manual effort in comparison to more traditional techniques, such as manual laser distance meter or clinometer. However, other approaches are also possible, such as UAV lidar.
With airborne surveys, the detection of top canopy is not affected by occlusion issues, but the capability to precisely hit the tree tops and map the ground level is influenced by various factors, including forest structure, lidar pulse density, scan angle, platform altitude, and beam size, among others. Different studies reported underestimation of tree heights with ALS, especially in old growth forests.
Low cost UAVs systems equipped with a consumer-grade RGB camera can acquire stereo imagery, from which a high resolution point cloud can be generated and processed to obtain a DSM. Using a local maxima algorithm the identification of tree tops is possible, and then by means of ground level subtraction, tree heights can be estimated. However, interpolating the terrain with only the few points obtained by stereo optical imagery, which differently from lidar does not penetrate the canopy, can be a problem (Rosca et al., 2018). A high resolution Digital Elevation Model (DEM) obtained from other sources is therefore necessary to estimate individual tree level metrics. When ground information is available, UAVs can represent a cost-effective approach. Only limited information is available on the accuracy of tree heights as measured by different techniques or sensors over the same forest. Considering the high economic costs of forest surveys, independently by the adopted instrument or technique, a comparative analysis over a single site can be useful, as it provides valuable information to help decision making in forest resources management and planning. The objective of this research is to comparatively analyze tree heights as derived by: a ground-based traditional technique, stereo-photogrammetry with imagery from a UAV, and ALS. Differences in measured heights as well as their impact on above ground biomass estimation were analyzed. The research was realized in a dense tropical forest in Ghana, in the Ankasa Conservation Area, where occlusion problems are relevant.
Photogrammetric data were collected in March 2016 using a Phantom 3 Professional UAV equipped with a 4K camera. The Pix4Dcapture software (Pix4D) guided the flight plan: 215 RGB images were collected at 5 cm spatial resolution. Using Agisoft Photoscan Professional software (Agisoft LLC) a point cloud and orthophotos were produced. Airborne lidar (ALS) data were collected in March 2012 with a Optech GEMINI sensor, collecting up to 4 range measurements. Mean laser point density ranged between 12 and 20 points per square meter. The point cloud was classified using Terrascan software (Terrasolid) into ground and vegetation returns. A botanical survey was realized in March 2016 collecting species, diameter at breast height (DBH), and height information for trees having DBH above or equal to 10 cm. The height of the trees was measured with a transponder and vertex meter. Wood density values at species level were extracted from the Global Wood Density Database. Above ground biomass (AGB) was computed according to Chave et al. (2014) equation.
A total of 329 trees belonging to 82 different species with DBH > 10 cm were recorded in the botanical survey, with a recorded maximum height of 55.5 m and a mean height equal to 20.6 m. The co-registration between ALS point cloud (LPC) and photogrammetric point cloud (PPC) was characterized by a mean point distance of 0.44 m, a standard deviation equal to 0.52 m, and root mean squared error (RMSE) of 0.62 m.
The photogrammetry surveys was realized in 2016, instead, ALS survey was done in 2012: the derived DSMs shows differences due to changes occurred in the canopy along time.
The comparison of CHMs at 1 m spatial resolution showed lower maximum and mean height values for photogrammetry data with respect to ALS. Similarly, the crown detection with Inverted Watershed Analysis (IWA) produced much lower amount of individuals detected when using photogrammetric data, with respect to ALS. The histogram of tree heights from photogrammetric data showed not only less individuals but also a different distribution in heights compared to the other histograms. These results first indicate that the number of trees that can be detected with IWA procedure is much lower than the trees recorded by ground surveys, as expected. Ground surveys, either with traditional techniques or TLS remains fundamental for full inventory purposes.
Thanks to co-registration of point clouds, the height values for 23 large trees (DBH > 30 cm) were extracted from ALS and photogrammetry datasets. For the 23 trees the agreement between the ALS and photogrammetric height was very good, with a coefficient of determination around 0.95; this agreement suggests that overestimation of tree height occurred when collecting data with ground traditional technique.
The total above ground biomass (AGB) for the 23 trees was computed using in input the DBHs collected in ground survey, the species-level wood density values, and heights according to ground, ALS and photogrammetric derived heights. The AGB estimated using ground height values was 6.3 %, and 4.5 % higher than that estimated using ALS and photogrammetric heights, respectively. If the ratio between AGB from 23 trees to that of the entire plot was assumed as constant, the biomass computed using ground heights resulted 9 %, and 6.5 % higher than that estimated using ALS and photogrammetric heights, respectively.
Several conclusions can be derived from this study. First, a limit to the use of photogrammetric technique for deriving tree individual metrics is represented by the availability of a high resolution DTM, that should be available or collected from other sources.
Findings from different previous studies evidence that the accuracy of height estimated in the ground or by ALS largely depends on specific site conditions, instruments settings, and the ability of the operators; however measuring tree heights of cone-shaped species can be easier from the ground, as the tree top can be more evident.
This and other research suggest that in dense forests, where sight occlusion problems occur, ground traditional techniques can lead to overestimation of tree heights, unless the tree tops are clearly visible and distinguishable, such as in the case of cone-shaped trees or open forests. On the other hand, with ALS and photogrammetry techniques underestimation can occur, in variable amount according to local and instruments conditions.
When enough resources are available, the ALS technique may represent an advantage, as it can cover relatively large areas in a reduced time, and allows the generation of a DEM. From photogrammetric data lower heights were obtained in this research, however the economic convenience of this techniques supports its use when a DEM is available and for repeated canopy monitoring. A promising new opportunity is provided by low flying drone-based lidar systems with wide viewing angles that also record point clouds looking side-ways into forests allowing estimations stem characteristics such as DBH (Brede et al., 2017).
Height is certainly responsible for a part of the uncertainty associated to AGB estimates: more accurate height measurements can help to reduce the uncertainty in AGB mapping efforts at any scale.


Brede, B., Lau, A., Bartholomeus, H. M., & Kooistra, L. (2017). Comparing RIEGL RiCOPTER UAV LiDAR derived canopy height and DBH with terrestrial LiDAR. Sensors, 17(10), 2371.

Chave, J., Réjou-Méchain, M., Búrquez, A., Chidumayo, E., Colgan, M. S., Delitti, W. B., ... & Henry, M. (2014). Improved allometric models to estimate the aboveground biomass of tropical trees. Global change biology, 20(10), 3177-3190.

Roşca, Sabina ; Suomalainen, Juha ; Bartholomeus, Harm ; Herold, Martin (2018). Comparing terrestrial laser scanning and unmanned aerial vehicle structure from motion to assess top of canopy structure in tropical forests, Interface Focus 8 (2).

The latent heat flux (LE) plays an important role in surface-atmosphere interactions. Traditionally, LE or the actual evapotranspiration (Eta) is measured by means of lysimeters, eddy-covariance systems or fiber optics, providing estimates which are spatially restricted to a footprint from a few square meters up to several hectares. In the past, several methods have been developed to derive LE by means of multi-spectral remote sensing data using thermal and VIS/NIR satellite imagery of the land surface. As such approaches do have their justification on coarser scales, they do not provide Eta information on the fine resolution plant level over large areas which is mandatory for the detection of water stress or tree mortality. In addition, the validation of such approaches is often difficult, as the inherent coarse resolution of the available satellite sensors is often much more larger than the contributing footprint of the in-situ measurements.
In this study, we present a comparison of a drone based assessment of LE with eddy-covariance measurements over two different forest types - a deciduous forest in Alberta, Canada and a tropical dry forest in Costa Rica. Drone based estimates of LE were calculated applying the Triangle-Method proposed by Jiang and Islam (1999). In our study, the Triangle-Method estimates LE by means of the Normalized Difference Vegetation Index (NDVI) and land surface temperature (LST) provided by two camera systems (MicaSense RedEdge, FLIR TAU2 640) flown simultaneously on an octocopter.
Results indicate a high transferability of the original approach from Jiang and Islam (1999) developed for coarse to medium resolution satellite imagery to the high resolution drone data, leading to a deviation in LE estimates of 10% compared to the eddy-covariance measurements. In addition, the spatial footprint of the eddy-covariance measurement can be detected with this approach, by showing the spatial heterogeneities of LE and thus Eta due to the spatial distribution of different trees and understory vegetation.

Motivation

Wildlife censuses play an increasing role in stemming extinctions of large mammal species [1, 2]. Obtaining statistics of the number of species, herd sizes and grazing locations are vital aids for game reserve managers to take actions against threats like illegal livestock poaching. In this context, recent years have brought a shift from manned surveys to (semi-) automated censuses involving Unmanned Aerial Vehicles (UAVs), which allow acquiring imagery of large areas at low costs and ensuring the safety of the operators [3].

Particular successes have been achieved in combining UAVs with automatic object detectors from computer vision [4]: state-of-the-art models employ deep Convolutional Neural Networks (CNNs; [5]) and have recently proven their ability to locate sparse animals in many large images with impressive precision [4]. However, such models fall short when it comes to repeated data acquisitions, which are common means to monitor mammal population dynamics over time. By that, we mean the re-application of an existing model to a new region or a new data acquisition in a later moment. Inherent discrepancies between datasets, known as domain shifts, have the effect that animal detectors trained on one dataset cannot directly be applied to another one. This generic problem, often observed in data science, is a real problem in remote sensing data analysis and is being considered extensively in recent literature [6].

In this work, we tackle this problem by building on a crucial observation: human operators are primarily interested in locating the animals in new UAV acquisitions, much more than in obtaining a well-performing automated animal detector model. As such, instead of obtaining labels in an expensive campaign and re-training an entire model for a new dataset, it can be sufficient to re-use an existing model and rank the predictions in such a way that the true positives (animals) are found first. In the following, we thus present a ranking strategy that allows to efficiently retrieve instances in the new image (or target domain) and separate them from false detections.




Proposed Method

Figure 1 shows an overview of our proposed animal detection heuristic. At the core we employ the deep CNN- based animal detector presented in [4], trained on the source dataset. We use it to predict potential animal locations (candidates) in both the source and target datasets, and for every candidate we store its feature vector as the output from the second-to-last layer of the CNN.

The two point clouds in Figure 1 (“predicted samples”) show the source (top) and target (bottom) candidate feature vectors projected using t-SNE [7]. As can be seen, the two datasets show strong differences in the number of correctly predicted animals (blue) and false alarms (red); obviously, the source model is unable to obtain high precision on the target dataset. To find the animals quickly, we thus rank the target predictions according to their “likelihood” of being an animal. To this end, we exploit the model’s superior performance in the source dataset as follows:

1. Rank the source candidates by means of an auxiliary classifier confidence. In detail, we fit an SVM with polynomial kernel (degree 2) to the candidates and record the samples’ distance to the hyperplane.

2. The source sample distances are then transferred to the target samples that lie at a similar location in the feature space. To do so, we resort to Optimal Transport (OT; [8]). OT is a family of domain adaptation methods that establish a transport plan, which minimizes the individual costs between samples while respecting the overall transported “mass”. This allows to identify specific source-to-target matches, which can be used to transfer the source rank scores to the target samples.

3. Re-rank the target samples according to the transported source costs and process them in order: for every selected sample, we crop a window around it in space to include as many other candidates as possible and record the number of animals found therein.




Experiments and Results

We evaluate our selection strategy on two UAV acquisitions gathered over the private-owned Kuzikus wildlife reserve, Namibia. We use 450 images of size 4000 × 3000 with 830 animals as a source dataset and 2841 images of larger sizes containing 576 animals for our target dataset. We train our detector CNN on the source dataset and then use it to find candidates for both. We iteratively draw 1000 samples and record the number of animals found according to our sample selection strategy and compare it to three competing sample selection strategies (random selection, Breaking Ties [9], ordering by animal confidence of the source CNN).

Figure 2 shows the number of animals found in the target dataset for the four different strategies. As expected, random sampling and Breaking Ties fail at finding the animals in the first iterations, because of the large class imbalance and the fact that Breaking ties goes for the most borderline examples, respectively. However, we note that our proposed scheme (red) is consistently better than the others, even when using the CNN’s prediction confidence (green).




Conclusion

We addressed the problem of re-using an object detector (a deep CNN), trained to detect animals on one UAV- derived image dataset, on another. Different properties between the two datasets cause machine learning models to fail. In our context of animal censuses, we can however exploit the fact that model precision is only of minor interest compared to recall, i.e. the ability to locate all the animals in the images. We thus present a strategy that uses a trained CNN to transfer its proficient prediction from the source domain to the new target dataset. To this end, we make use of Optimal Transport, which establishes globally distribution-aware source-to-target sample correspondences. Experiments show that our strategy is particularly effective in prioritizing true animals (instead of false positives), when compared to conventional sampling heuristics. It therefore allows for a more efficient localization of animals without the need of acquiring expensive annotations for retraining the model in the target domain.




References

[1] L. Wilson-Wilde, “Wildlife crime: A global problem,” Forensic Sci. Med. Pathol., vol. 6, no. 3, pp. 221–222, 2010.

[2] E. Bondi, F. Fang, M. Hamilton, D. Kar, D. Dmello, J. Choi, R. Hannaford, A. Iyer, L. Joppa, M. Tambe, and Others, “Spot poachers in action: Augmenting conservation drones with automatic detection in near real time,” 2018.

[3] J. Linchant, J. Lisein, J. Semeki, P. Lejeune, and C. Vermeulen, “Are unmanned aircraft systems (UASs) the future of wildlife monitoring? A review of accomplishments and challenges,” Mamm. Rev., vol. 45, no. 4, pp. 239–252, 2015.

[4] B. Kellenberger, D. Marcos, and D. Tuia, “Detecting mammals in UAV images : Best practices to address a substantially imbalanced dataset with deep learning,” Remote Sens. Environ., vol. 216, no. December 2017, pp. 139–153, 2018.

[5] A. Krizhevsky, I. Sulskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” Adv. Neural Inf. Process. Syst., pp. 1–9, 2012.

[6] D. Tuia, C. Persello, and L. Bruzzone, “Recent advances in domain adaptation for the classification of remote sensing data,” IEEE Geosci. Remote Sens. Mag., vol. 4, no. 2, pp. 41–57, 2016.

[7] L. J. P. Van Der Maaten and G. E. Hinton, “Visualizing high-dimensional data using t-SNE,” J. Mach. Learn. Res., vol. 9, pp. 2579–2605, 2008.

[8] N. Courty, R. Flamary, D. Tuia, and A. Rakotomamonjy, “Optimal Transport for Domain Adaptation,” vol. X, no. X, pp. 1–14, 2015.

[9] T. Luo, K. Kramer, D. B. Goldgof, L. O. Hall, S. Samson, A. Remsen, and T. Hopkins, “Active learning to recognize multiple types of plankton,” J. Mach. Learn. Res., vol. 6, pp. 589–613, 2005.