The challenge of classifying and locating Phoenix palm trees in different scenes with different appearances and varied ages has been addressed with deep learning object detection over aerial images. Nevertheless, an explicit limitation hereof is that palms should be visually identifiable in the image—i.e., palm crowns should be larger than the pixel size. Unfortunately, high-spatial resolution imagery is not widely and directly available in the Phoenix palm growing regions of the Mediterranean, Middle East, and North Africa. This study, therefore, presents the re-implementation of a semantic segmentation architecture to train a model able to classify Phoenix palm pixels. This is applied to freely available medium resolution space-borne Sentinel-2 images over the Spanish island of La Gomera (Canary Islands). At the study site, a total of 116,330 Phoenix palms had been inventoried by the local government. Palms appear in multiple, heterogeneous environments, which implies a background variability that is a persistent challenge for palm pixel classification. The re-implemented architecture is a novelty in deep semantic segmentation and density estimation initially developed for counting objects of sub-pixel size. And it proved to be successful for creating a model of palm classification, thereby compensating for the limited spatial resolution of the Sentinel-2 images. The palm tree sub-pixel classification model achieved an overall accuracy of 0.921, with a recall and precision of 0.438 and 0.522. These results demonstrate the potential of remote sensing data of medium-spatial resolution for vegetation mapping in applications where trees are scattered over extensive areas.
Remote sensing based land cover classification in urban areas generally requires the use of subpixel classification algorithms to take into account the high spatial heterogeneity. These spectral unmixing techniques often rely on spectral libraries, i.e. collections of pure material spectra (endmembers, EM), which ideally cover the large EM variability typically present in urban scenes. Despite the advent of several (semi-) automated EM detection algorithms, the collection of such image-specific libraries remains a tedious and time-consuming task. As an alternative, we suggest the use of a generic urban EM library, containing material spectra under varying conditions, acquired from different locations and sensors. This approach requires an efficient EM selection technique, capable of only selecting those spectra relevant for a specific image. In this paper, we evaluate and compare the potential of different existing library pruning algorithms (Iterative Endmember Selection and MUSIC) using simulated hyperspectral (APEX) data of the Brussels metropolitan area. In addition, we develop a new hybrid EM selection method which is shown to be highly efficient in dealing with both imagespecific and generic libraries, subsequently yielding more robust land cover classification results compared to existing methods. Future research will include further optimization of the proposed algorithm and additional tests on both simulated and real hyperspectral data.
High spatial resolution satellite imagery provides an alternative for time consuming and labor intensive in situ measurements of biophysical variables, such as chlorophyll and water content. However, despite the high spatial resolution of current satellite sensors, mixtures of canopies and backgrounds will be present, hampering the estimation of biophysical variables. Traditional correction methodologies use spectral differences between canopies and backgrounds, but fail with spectrally similar canopies and backgrounds. In this study, the lack of a generic solution to reduce background effects is tackled. Through synthetic imagery, the mixture problem was demonstrated with regards to the estimation of biophysical variables. A correction method was proposed, rescaling vegetation indices based on the canopy cover fraction. Furthermore, the proposed method was compared to traditional background correction methodologies (i.e. soil-adjusted vegetation indices and signal unmixing) for different background scenarios. The results of a soil background scenario showed the inability of soil-adjusted vegetation indices to reduce background admixture effects, while signal unmixing and the proposed method removed background influences for chlorophyll (ΔR2 = ~0.3; ΔRMSE = ~1.6 μg/cm2) and water (ΔR2 = ~0.3; ΔRMSE = ~0.5 mg/cm2) related vegetation indices. For the weed background scenario, signal unmixing was unable to remove the background influences for chlorophyll content (ΔR2 = -0.1; ΔRMSE = -0.6 μg/cm 2 ), while the proposed correction method reduced background effects (ΔR2= 0.1; ΔRMSE = 0.4 μg/cm2). Overall, the proposed vegetation index correction method reduced the background influence irrespective of background type, making useful comparison between management blocks possible.
Spectral Mixture Analysis is a widely used image analysis tool with many applications. Yet, one of the major issues with
this technique remains the lack of ability to properly account for the spectral variability of endmembers or ground cover
components that occur throughout an image scene. Endmember variability is most often addressed using iterative
mixture cycles (e.g. MESMA) in which different endmember combination models are compared for each pixel. The
model with the best fit is assigned to the pixel. The drawback of MESMA is the computational burden which often
hampers the operational use. In an attempt to address this issue we proposed a new geometric based methodology to
more efficiently evaluate different endmember combinations in MESMA. This geometric unmixing methodology has a
two-fold benefit. First of all, geometric unmixing allows a fast and fully constrained unmixing, which was previously
unfeasible in MESMA due to the long processing times of the available fully constrained unmixing methods. Secondly,
whereas the traditional MESMA explores all different endmember combinations separately, and selects the most
appropriate combination as a final step, our approach selects the best endmember combination prior to unmixing, as such
increasing the computational efficiency of MESMA. To do so, we built upon the equivalence between the reconstruction
error in least-squares unmixing and spectral angle minimization in geometric unmixing. With the inclusion of the
proposed endmember combination selection technique, the computation time decreased by a factor between 5 and 8.5,
depending on the size and organization of the libraries. The spectral angle can as such be used as a proxy for model fit,
enabling the selection of the proper endmember combination from large spectral libraries prior to unmixing.
Alternating Least Squares (ALS) is a blind source separation method commonly used in Chemometrics to simultaneously estimate the absorption spectrum and concentration of the different components in a chemical sample. In this study, the transferability of ALS from Chemometrics to agricultural Remote Sensing is evaluated. Due to the subpixel contribution of background components, spectral unmixing has become an indispensable processing step in the spectral analysis of agricultural areas. Yet, traditional unmixing techniques only allow estimating the sub-pixel cover distribution of the different components, but fail to provide an estimate of the pure spectral signature of the crop. This info is, however, highly valuable as this pure crop signature could be used to monitor the health status of the trees. Here, we anticipate that ALS can provide a solution. ALS estimates both the concentration and the absorption spectra of the different components in a chemical sample and this can easily be translated into estimating both the subpixel cover fraction and spectral signature of the different components in a mixed image pixel. The ALS model was tested on simulated hyperspectral images of Citrus orchards in which ray-tracing software was used to realistically incorporate spectral variability, multiple scattering and shadowing effects. Both the accuracy of the extracted cover fractions and the pure spectral signatures of the crop were assessed, as well as the accuracy with which the biophysical parameters of the trees (i.e. chlorophyll content, leaf water content and Leaf Area Index) could be derived from the extracted crop signature. ALS indeed allowed to simultaneously estimate the subpixel cover distribution (RMSE = 0.05), as well as the pure spectral signatures of the different endmembers (RRMSE < 0.12), and considerably improved the extraction of biophysical parameters (ΔR2 up to 0.43). ALS thus provides a promising new image analysis tool for agricultural remote sensing.
The use of imaging spectroscopy for florisic mapping of forests is complicated by the spectral similarity among coexisting
species. Here we evaluated an alternative spectral unmixing strategy combining a time series of EO-1 Hyperion
images and an automated feature selection strategy in MESMA. Instead of using the same spectral subset to unmix each
image pixel, our modified approach allowed the spectral subsets to vary on a per pixel basis such that each pixel is
evaluated using a spectral subset tuned towards maximal separability of its specific endmember class combination or
species mixture. The potential of the new approach for floristic mapping of tree species in Hawaiian rainforests was
quantitatively demonstrated using both simulated and actual hyperspectral image time-series. With a Cohen’s Kappa
coefficient of 0.65, our approach provided a more accurate tree species map compared to MESMA (Kappa = 0.54). In
addition, by the selection of spectral subsets our approach was about 90% faster than MESMA. The flexible or adaptive
use of band sets in spectral unmixing as such provides an interesting avenue to address spectral similarities in complex
vegetation canopies.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.