
Table of Contents
-
Introduction i. Invasive Plant Species ii. Satellite Mission iii. Synthetic Aperture Radar iv. Multi-Spectral Instruments v. Algorithms
-
Method i. Mapping of Invasive Species ii. Tree Population in Mixed Forests
-
Results i. Mapping of Invasive Species ii. Tree Population in Mixed Forests
-
[Outlook](#52a2) & [Conclusion](#a4ab) i. Conclusion ii. Outlook
Results
Mapping of Invasive Species
In Kattenborn’s experiment using UAVs, hyperspectral data offered the highest accuracy for three of the four tested invasive species. The fourth instead got similar results when using the hyperspectral data and the RGB data In general, the accuracy increased by combining different predictors. The best combination was hyperspectral, texture, and 3D structure, which offered the highest accuracy. When comparing the satellite data, Sentinel-2 with its multispectral imagery offered more value to the fraction of vegetation cover than Sentinel-1 with its SAR data [1].
In all three papers, Satellite Imagery was used for data collection, while Kattenborn further improved the training process and decided to not just rely on field surveys or hand-labeled satellite data but create a ground truth using UAVs. When we look at the classification outcome Forstmaier achieved, prediction accuracy of 92.5% was the result. Nevertheless, it is noteworthy that sensitivity and specificity highly differ in the observed region. Furthermore, the created binary map doesn’t represent smaller eucalyptus patches good enough since there has to be found a compromise when choosing the threshold value [2]. Deng and his team managed to get a slightly better result and got an overall classification accuracy of 93.59%. Even though two different sources were used to build a ground truth, the authors worry that the presented approach is underestimating the eucalyptus plantations caused by a lack of data. As reference data, a field survey in 2017 and high-resolution Google Earth data were considered. In addition to the classification map, the team was also able to create a calendar when eucalyptus trees were planted partitioned on a yearly basis, thanks to the gathered data [3].
![Figure 2.1: Map of the Iberian peninsula showing the occurrence of tree plantations. The map was created with supervised classification and manual polygon delineation of Landsat, SPOT, or RapidEye satellite imagery. Image dates vary but can be centered around 2015. The resolution varies too and depends on each country [6].](https://towardsdatascience.com/wp-content/uploads/2021/07/1cAxgtGMQu23ruvuvPAtJ_w.png)
When looking at the tree plantation map of the Iberian peninsula, a strong correlation with the eucalyptus occurrence can be found. Globalforestwatch offers forest image data free of charge and different data layers can be applied while choosing the area of interest. Thus, a map showing the distribution of managed tree plantations on the Iberian peninsula could be created.
Fig. 2.1 proves that the majority of the managed tree plantations on the Iberian peninsula are located in the west and are used for wood fiber and timber. By comparing Fig. 2.1 with Fig. 2.2, one can see that this result overlaps with the outcome Forstmaier et al. showed in [2], mapping the eucalyptus distribution on the Western Iberian peninsula.
![Figure 2.2: Map showing the eucalyptus occurrence in the Western Iberian peninsula in percentage with respect to the maximum forest cover (FC) between 2010 and 2018 [7].](https://towardsdatascience.com/wp-content/uploads/2021/07/184nyMAPdMkKfnanD95tP9A.png)
Tree Populations in Mixed Forests
When we compare the results of the tree species classification approaches, one can find noticeable differences. While Persson et al. managed to get an overall classification accuracy of 88.2%, Immitzer and his team achieved an overall accuracy of 82.4%. Both teams used the whole available dataset for their best outcome and also all of the available spectral bands. They both achieved good performance when using fewer input data in form of less imagery and less used spectral bands, but came to the same conclusion that each additional band improves the outcome, even though just slightly. Persson et al. pointed out that an image taken in May delivered the best result when relying on just a single image (overall accuracy of 80.5%). The most important bands are the 3 red edge and NIR bands. Tree species on provided imagery using such bands are already highly separable [4]. On the other hand, Immitzer et al. achieved an excellent result when it comes to distinguishing between broad-leafed and needle-leafed trees with not a single misclassification. Also, using the 4 standard bands instead of all 8 leads to a just slightly worse accuracy. While the result in this study highly depends on the observed tree species, different other studies underline this statement and point out that the number of tree species and the tree species themselves play a major role. Furthermore, the authors recommend using an object-based classification approach since the pixel-based approach provides noteworthy lower accuracy. The most important bands turn out to be Green, Near-Infrared 1, and Blue.
Also, using visual photo interpretation to create an adequate image data set for verification is advisable, since field surveys (i.e. measurements from the ground) can provide another outcome than a top-down perspective. The overall classification accuracy for 10 trees was around 82% (object-based) and the producer’s accuracy between 33% (European hornbeam) and 94% (European beech). Object-based approaches provide better results than pixel-based ones and the classification accuracy increases when the full spectral resolution was used instead of a subset [5].
In summary, both remote sensing tools (Sentinel-2 and WorldView-2) are highly suitable for providing imagery regarding a classification task of tree species in mixed forests, thanks to their spatial and spectral properties.
Fig. 2.3 shows biodiversity hotspots in Southern Europe. Conservation Internationals biodiversity hotspots are marked with red and are more often located in Southern Europe. The dataset was collected in 2011 and only covers endangered areas on land. Following criteria must be met for Conservation Internationals biodiversity hotspots:
- At least 1500 species of vascular plants (larger than 0.5% of the world’s total) are endemic.
- At least 70% of the original natural vegetation has been lost.
On the other hand, green mark areas of tree cover with a resolution of 30m 30m resolution. The content was captured in the years 2000 and 2010, using Landsat-7 and MSI.
![Figure 2.3: Map with biodiversity hotspots in Central and Southern Europe. Red areas mark regions with low biodiversity because of endemism and human threat while green areas mark tree cover density [8].](https://towardsdatascience.com/wp-content/uploads/2021/07/1lTOeeuPLtgrbHfLKUmG7tw.png)
Outlook and Conclusion
Conclusion
All in all, each paper underlines the importance of monitoring invasive species. The approach to gathering imagery with UAVs offers not only a way to get reference data for satellite classification faster and less cost-intensive, but also with higher quality. This leads to a more accurate satellite classification map. Combining different information layers such as spectral, textural, and structural results in an increased map accuracy. Nevertheless, the mapping quality also depends on the plant species and its properties (e.g. plant occurring in smaller patches) [1].
Detection of invasive species like eucalyptus and monitoring them is crucial to keep the biodiversity of a region high. By doing so, sensitivity and specificity do strongly depend on the chosen parameters. Even though FNN is a fairly simple approach, the neural network is able to solve different landcover classification tasks well. The classification accuracy is high, but still not without errors. Thus, applying FNN on multispectral imagery provided by Sentinel-2 proves to be a suitable method for monitoring invasive species. In addition, the study pointed out that nine Natura 2000 areas (protected areas) in Portugal are heavily affected by the eucalyptus tree [2].
A downside, satellite data often has is its dependence on the weather. Frequent clouds can cause a lack of data and since satellites orbit certain parts of the earth only once in a while, it’s difficult to compensate for this loss. The Sentinel-2 mission lowers this risk since the repeating rate is higher. Also, Synthetic Aperture Radar (SAR) is another way to reduce the dependency on weather, because this technology is less impaired by clouds using a different wavelength [3].
Short rotation plantation is a growing phenomenon in different areas of the world. Some scientists draw attention to the possible dangers for the ecosystem, but a large-scale observation of these plantations is necessary to get further insights into the impact.
When we look at the comparison of mapping Eucalyptus using satellites, Deng and Forstmaier were able to create maps with significant accuracy. Even though both teams did not use the same test set, Forstmaier et al. managed to get almost the same outcome as Deng, while using a significantly less complicated approach with fewer input data.
By looking at the classification of tree populations in mixed forests both teams used multispectral satellite imagery. Persson et al. achieved an overall accuracy of 88.2% using all bands and each of the 4 images. The most important image turns out to be taken in May. It offers an overall classification accuracy of 80.5% when not considering the other 3. Then, adding the other months increased the accuracy successively. Using just 13 of all 40 wavelengths decreased the result slightly (overall accuracy of 86.3%) and the most important bands are red-edge, narrow NIR band, and most SWIR bands [4]. The team around Immitzer achieved an overall accuracy of 82% when classifying 10 different tree species with 8 spectral bands [5].
Some tree species had a significantly lower classification accuracy compared to others and the authors suggest increasing temporal resolution to improve the classification.
Still, one needs to consider the important differences when comparing both attempts of tree species classification, which finally lead to the deviation of accuracy. Both teams examined different tree species and a different number of tree species which highly impacts the overall result. But they also used different satellite missions. The WorldView-2 mission is not only 6 years older than Sentinel-2 but also the camera systems differ, respectively.
Outlook
Even though the approaches were different, they share the same goal: to improve the mapping, especially of invasive species, to understand their impact on the environment, and finally to protect humans, animals, and other native plants. While several countries, especially islands, prohibited the import of foreign species already a long time ago, other nations tried to exploit certain exotic plants. Now they have to deal with some serious threats to the native environment while the total extent is still unclear.
Different approaches were presented to improve the observation procedure. The semi-automated way of collecting reference data is the first step in the right direction. Of course, UAVs can not cover as large areas as satellite imagery does. But training and validation data are essential for improving the outcome of a classification task.
On the other hand, the current limits can be found in the satellite imagery itself. If data with a higher resolution would be provided to researchers, even better results could be achieved while reducing the effort. More satellites would provide an increased temporal resolution while better instruments could increase the spatial resolution. In the future, ESA is extending the Earth Observation mission by bringing a few more satellites into orbit, carrying different tools which hopefully improve the surveillance of vegetation too.
To further improve the FNN approach suggested by Forstmaier et al., the input data layer could be expanded. Instead of just using the Sentinel-2 and image data collected by the MSI, data provided by other missions and different remote-sensing tools could be combined.
Relying on a large temporal resolution or even expanding that seems to be problematic when it comes to mapping vegetation or even invasive plant species. Even though this is applicable for large-scale rotation plantations, natural occurrence patterns change regularly. Thus, data from a previous year can differ significantly from the actual one. Also, there is a significant delay when it comes to collecting a dataset with a large temporal resolution. Data provided by newly introduced remote-sensing tools cannot be used immediately but scientists rather would need to wait until a reasonable number of months or years passed.
LiDAR could be a different data layer to further improve tree classification, especially when it comes to reference data collection. LiDAR measurement equipment is getting more and more reliable since manufactures start using MEMS technology while the prices decrease significantly. LiDAR data collected by UAVs could not only offer data of tree crowns but also unveil underlying vegetation.
Continue with part 1 or part 2.
Bibliography
[1] Teja Kattenborn, Javier Lopatin, Michael F orster, Andreas Christian Braun, and Fabian Ewald Fassnacht. UAV data as alternative to field sampling to map woody invasive species based on combined Sentinel-1 and Sentinel-2 data. Remote Sensing of Environment, 227:61–73, 2019.
[2] Andreas Forstmaier, Ankit Shekhar, and Jia Chen. Mapping of Eucalyptus in Natura 2000 Areas Using Sentinel 2 Imagery and Artificial Neural Networks. Remote Sensing, 12(14), 2020.
[3] Xinping Deng, Shanxin Guo, Luyi Sun, and Jinsong Chen. Identification of short-rotation eucalyptus plantation at large scale using multi-satellite imageries and cloud computing platform. Remote Sensing, 12(13), 2020.
[4] Magnus Persson, Eva Lindberg, and Heather Reese. Tree species classification with multitemporal sentinel-2 data. Remote Sensing, 10(11), 2018.
[5] Markus Immitzer, Clement Atzberger, and Tatjana Koukal. Tree species classification with random forest using very high spatial resolution 8-band worldview-2 satellite data. Remote Sensing, 4(9):2661–2693, 2012.
[6] Screenshot taken by the author: Map of tree plantations on the Iberian peninsula. [Online; accessed March 15, 2021], URL: https://www.globalforestwatch.org/map/.
[7] Composition created by the author, using images provided by Forstmaier et al. in [2]. Eucalyptus distribution on the western Iberian peninsula.
[8] Screenshot taken by the author: Map of Europe showing regions of low biodiversity. [Online; accessed March 15, 2021], URL: https://www.globalforestwatch.org/map/.
[9] Screenshot taken by the author: Housten Texas, SAR, 2019. [Online; accessed March 10, 2021], URL: https://www.esa.int/Applications/Observing_the_Earth/Copernicus/Sentinel-1.
[10] Screenshot taken by the author: Kyiv Ukraine, MSI, 2020. [Online; accessed March 10, 2021], URL: https://www.esa.int/Applications/Observing_the_Earth/Copernicus/Sentinel-2.