FINAL TECHNICAL REPORT
Project: Monitoring dynamics of Croatian olive fields using remote sensing applications
Anita Simic1, Lorenzo Seguini2, Ante Graovac3,
Ivica Ljubenkov4, and Jing Chen1
1University of Toronto, Department of Geography, Canada
2European Commission, Ispra, Italy
3The Rudjer Boskovic Institute, Croatia
4SMS Prehrambeno Razvojni Centar d.o.o., Croatia
Prepared for the Ministry of Science, Education and Sports, Croatia –
Unity through Knowledge Fund (UKF) – Connectivity Program
Project duration: June 4th, 2009 - January 10th, 2010
Prepared by Dr. Anita Simic
The European Economic Community has placed regulations concerning the management of the olive oil market sector since 1975. Counting olive trees is crucial for the European Union’s olive production subsidies. This kind of statistical survey is important for the implementation of registers of olive fields in Croatia, which are the basis for the crop management within the frame of the Common Agricultural policy of the European Union. Up until now the accounting of olive trees in Croatia is generally based on data submitted by the plantation owners.
Remote sensing is a valuable tool for agricultural statistics. Quality and quantity of both wine and olive products will be regulated in the coming years when Croatia is expected to become a member of the European Union and when the export of these products becomes an important part of the Croatian economy.
If the remote sensing technology is adapted by Croatian scientists, future statistical survey of crops and fruit trees in Croatia will be cost-efficient and more effective. This project enhances future counting and monitoring of olive fields and vineyards across the whole country using remote sensing technology. This project aims to provide the assessment of number of olive trees in the Marina region in Dalmatia.
Our intention is also to introduce remote sensing as a technology that can be applied to other fields as well. The results generated at this project stage will be used in the subsequent phases, where we will explore the quality of the olive fields and vineyards using hyperspectral remote sensing (Hyperspectral Remote Sensing Applications in Monitoring Dynamics of Croatian Vineyards and Olive Fields). Biochemical and biophysical parameters of fruit crops are crucial indicators of their health, fruit quality and yield. Biochemical constituents are, in particular, important indicators of chlorosis due to plant stress and nutritional deficiencies. A number of studies have indicated the feasibility of chlorosis detection through chlorophyll content estimation using either spectroscopy or optical indices proposed to relate remote sensing data to crop status.
SITE DISCRIPTION and FIELDWORK
The study site covers approximately 5x5 km area and it is situated in the Marina region in Dalmatia, Croatia. This region is generally flat with some small hills and it is highly populated with olive trees. Some areas are cultivated and olive trees are structured in rows; however, a large portion of the region is uncultivated with randomly distributed olive trees. In cultivated areas the background consists of soil and small pebbles, while uncultivated areas enclose mainly understory vegetation such as grass and shrubs. The size of trees varies from small and young to large and old olive trees. During the fieldwork (Jun 15-16 and July 01-09 / 2009) (Figure 1) Dr. Ljubenkov and Dr. Simic visited ‘Obrt Bilaja’ and Uljara located within the region. In June, several places were chosen and covered with white plastic cloths (2x2 m). The GPS instrument was used to measure geographic coordinates of these locations, which were later recognized on the images and used as the geo-reference points. In July, the fieldwork included measurements of structural parameters of the trees including height, crown radius, and tree geometry. The ASD instrument was also used to measure reflectance of leaves and background in order to enhance the ability of software to differentiate reflectance of the background from reflectance of trees.
Figure 1: Uncultivated and cultivated areas that are populated with olive trees in the region of Marina. Note: understory vegetation within an uncultivated area (bottom left) and soil background within a cultivated area (bottom right).
DATA and SOFWARE
High spatial resolution multispectral imagery (QuickBird) was used to identify and individually position olive trees within the area. In accordance with the European Union (EU), we used the Very High Resolution (VHR) satellite images (QuickBird). QuickBird images (Digital Globe) in Panchro and Pansharp mode (0.6 m pixel resolution) were acquired over the Marina region in early July 2009. Designed by the Joint Research Centre (JRC), a semi-automatic counting tool OLICOUNT was used for counting olive trees. This process is considered as an object-oriented type of classification. OLICOUNT operates with four parameters: grey value threshold, tree diameter, crown shape, and crown compactness. For reprojection of images the ENVI (Research Systems Inc.) software was used. ENVI was also used to create polygons used as parcels in OLICOUNT. The parcels were chosen based on the similarity between shape and size of olive trees as well as on the contrast in reflectivity between the background and trees.
Reprojection of the imagery into UTM coordinate system was necessary to visualize them into OLICOUNT. This work was mainly performed together with Mr. Lorenzo Seguini (European Commission, ISPRA, Italy). As OLICOUNT can handle a small area size, the whole region was divided in nine sub-areas of the same size using ENVI. As suggested by Mr. Seguini, each area was then separated into polygons (parcels) that were afterwards used in OLICOUNT. Based on the visual perception, the parcel polygons were created in ENVI and saved as shape files. Each layer of a shape file included several polygons of similar structural and reflectance characteristics.
Introduced to OLICOUNT, the parcels for a given area were laid over the appropriate image sub-areas. The ‘training’ window in OLICOUNT was used to visually delineate the olive trees for a given parcel layer within each area. Each parcel layer was then separately run. After each run, the visual inspection was performed. Additional manual corrections of input parameters were done when the delineation overestimated or underestimated the number of olive trees.
The total number of the trees was calculated as the sum of the trees in nine areas. The field validation after the classification has not been performed within the scope of this project.
RESULTS and DISSCUSSION
As explained in Methods, the image is divided in nine sub-images. We use one sub-area to explain the results and uncertainties within the classification; however, the last figure shows the final results for all nine sub-areas.
Panchromatic (black and white) image of Area 9 is shown in Figure 2. The most accurate results are observed within the cultivated areas, as the contrast between the trees and the background enhances the classification and delineation of trees (Figure 2). Olive trees have distinguished reflectance; the central part of a tree has commonly somewhat lighter spectrum likely due to the structural characteristics as a result of tree pruning. It is important to observe the shade seen on the left side of the tree (Figure 2), which in some cases can cause confusion with reflectance from the tree or background. Therefore, every analysis is unique and the Sun-sensor-target geometry and, thus, location of tree shades, plays an important role in the classification and tree counting.
Figure 2: The satellite panchromatic subset image (top figure) that includes a cultivated area (bottom left). Note the reflectance of an olive tree (bottom right) with a different reflection in the middle due to the structural characteristic of the trees as a result of tree pruning. The sub-setting of the images was done in ENVI.
Parcels (polygons) are chosen based on the similarity of tree structures and reflectance (Figure 3). This process is subjective based on the visual perception of an analyst. The performance of OLICOUNT is known as being successful for smaller cultivated areas and, uncultivated areas are commonly excluded from the process. However, random and uncultivated olive tree fields in Croatia are vastly spread and it is a challenge to use OLICOUNT for such tree distribution. Therefore, we separated polygons based mostly on the cultivation stage and differences in reflectance between trees and the background/understory. Figure 4 shows parcels in OLICOUNT. The software provides an option to select the polygons with similar characteristics based on the layers of a shape file; all selected polygons are shown in yellow using the same parameters during the classification (Figure 4).
Figure 3: Polygons (parcels) of an area based on the similar structural and reflectance characteristics. This vector (shape) file is created in ENVI.
Figure 4: Interface of OLICOUNT: The vector file overlaid over the image in OLICOUNT (top image); selected parcels with similar reflectance characteristics in OLIOCUNT (bottom image).
There are two ways how the structural parameters can be selected before the software perform the classification: (1) by visually selecting the tree in the Training window (Figure 5), or (2) by entering the parameters manually (Figure 8).
If the trees are selected visually, the software automatically calculates the main parameters (Figure 6); the parameters include min and max reflectance values and min and max tree crown radius.
Figure 5: Visual section of trees in the Training Window (OLICOUNT) (upper right image).
Figure 6: Training parameters created by the visual selection of trees in the training window.
After the calculations, OLICOUNT ‘recognizes’ the olive trees and places a dot over each tree (Figure 7). The chosen trees over the selected parcel polygons were than counted and the number is placed in the output text file.
Figure 7: Classification (recognition) of trees based on the visually chosen training sites (trees). Note that a small number of trees is not included in the final results.
During the visual inspection after the classification, a small number of trees is either omitted (not recognized) or committed (object other than trees are selected). These errors of omission and commission (Figure 8) are common for any type of classification and the goal of every classification is to minimize these errors.
As OLICOUNT provides an option of entering the input parameters manually (case (2)), we can manually adjust the values after the visual delineation of trees (case (1)) and rerun the classification. The right set up of the parameters can considerably reduce the errors (Figure 9).
Figure 8: Observed errors of omission and commission after the classification
Figure 9: Reconstruction of the input parameters performed to correct the classification. Note: In this case the classification is redone and based on the corrected parameters.
In addition to be able to enter the parameters manually and thus, to reduce the errors, after the classification is performed, each point can be manually deleted or added on the image. This is particularly important for very small trees that are often missed by OLICOUNT. This step further reduces the errors, although it is not commonly required (Figure 10).
As per conversation with Mr. Seguini from European Commission (ISPRA, Italy), OLICOUNT is ideal for cultivated areas and generally not used for the uncultivated areas. Although suggested by Mr. Seguini, as a possible option, to exclude these areas (Figure 11), our intention has been to try to incorporate both uncultivated and cultivated areas in our calculations. The Marina region is highly populated with randomly distributed trees within uncultivated areas, and to exclude these areas may result in underestimation of the counts.
Figure 11 demonstrates an uncultivated area where olive trees area randomly distributed. There are two main reasons for the results to be erroneous: (1) other vegetation, for instance, some other type of small trees, can be mistaken for small olive trees during the visual delineation; (2) reflectance between understory vegetation and olive trees is minimal; the software tends to produce high percentage of the omission or commission errors, where small olive trees are either missed, or, other type of vegetation is reported as olive trees, respectively. It is important to mention (per conversation with Dr. Ljubenkov) that a high precipitation rate in June enhanced the growth of understory vegetation within the region; thus, high greenness of understory provided similar reflectance as the overstory vegetation. Our future analysis should consider these factors and acquire images in different time of season when the reflectance contrast is more emphasised.
Figure 10: Manual correction can be used to add or delete trees
Figure 11: Uncultivated area populated with olive trees and other vegetation. Note: understory and trees have similar reflectance spectra
The best way to validate and correct the classification of the uncultivated areas is to perform the field validation. This step has not been part of this project; however, this or next projects of this type should include the field validation in order for the results to be considered as reliable. The process of validation includes a visit to the region of interest and comparing the vegetation in the field with the classified map. Additional information about the area, such as polygons based on the known vegetation types and percentage crop cover, could be incorporated in this project. Aerial photography could be also used to enhance the results.
Based on the above explained methods and results we develop the olive tree counts for all nine subareas (Figure 12). In total there are approximately 160,000 olive trees within the area of interest (approximately 5x5 km). We believe that the results are realistic (as per conversation with Dr. Ljubenkov).
Figure 12: Estimated number of olive trees within the whole image. Note: field validation is not part of this project and, thus, it is not performed.
This project phase aimed to provide the assessment of number of olive trees in one area of the Marina region in Dalmatia. Up until now the accounting of olive trees in Croatia has been generally based on the data submitted by the plantation owners. Using high spatial resolution multispectral imagery (QuickBird) and the OLICOUNT (European Commission) software, we were able to identify and individually position olive trees within the area. This technique may enhance the monitoring and counting of crops and fruit trees across the whole Croatia. The same technology has been widely accepted within the European Union.
This study phase will also enhance further development and use of remote sensing technology in Croatia, which will allow future monitoring of permanent crops across the whole country. The results generated within this project will be used in the subsequent phases, where we will explore the quality of the olive fields and vineyards using hyperspectral remote sensing (Hyperspectral Remote Sensing Applications in Monitoring Dynamics of Croatian Vineyards and Olive Fields).
Apan, A., Young , F., Phinn,m S., Held, A., and Favier, J., 2004. Mapping olive varieties and wihitn-field spatial variability using high resolution QuickBird imagery. Assessed from USQ ePrints http://eprints.usq.edu.au.
Bagli S. 2005. OLICOUNT Software Upgrade, Version 2.0. Technical Documentation
Masson, J. 2005. Use of very high resolution airborne and spaceborne imagery: a key role in the management of olive, nuts, and vineyard schemes in the frame of the common agricultural policy of the Europena Union. FRUTIC, Montpellier, France.
Zarco-Tejada, P.J., Berjon, A., Lopez-Lazano, R., Miller, J.R., Martin, P., Cachorro, V., Gonzales, M.R., de Frutos, A. 2005. Assessing vineyard condition with hyperspectral indices: Leaf and canopy reflectance simulations in a row-structured discontinuous canopy. Remote Sensing of Environment 99, 271-287.
--Ministry of Science, Education and Sports, Croatia for funding
--The Institute R. Boskovic for support and collaboration
--University of Toronto for in-kind contribution (instruments and ENVI software)
--European Commission for the OLICOUNT software
--Obrt Bilaja and Zajednica Maslinara I Uljara for their financial contributions