A Method to Count Olive Trees in Heterogenous Plantations from Aerial Photographs

Olive trees have been of economic and cultural value since pre-Roman times, and continue 1 to dominate landscapes and agriculture in many mediterranean regions. Recent mass losses of olive 2 trees in Southern Italy due to an exotic plant pathogen highlight the need for methods that to monitor 3 the olive trees in a landscape or region operationally. Here, we develop a method for counting olive 4 trees from aerial photographs and test it in areas with a high diversity of olive tree ages, sizes, and 5 shapes. This heterogeneity complicates tree counting as centennial trees often have crowns that 6 are split into multiple segments, resembling multiple crowns, while nearby crowns often form a 7 semi-closed canopy comprising multiple trees. Comparisons with reference counts in two 20 ha sites 8 and over three different years indicate the automated counts tend to be reasonably accurate (median 9 error 13%, n = 6), but heavily influenced by a few olive orchards with particularly high planting 10 densities and a relatively closed canopy in which distinguishing individual trees is challenging. 11 Overall, the algorithm estimated tree densities well (counting 82 to 109 trees/ha versus 87 to 104 12 trees/ha in the reference counts), indicating the method is suitable to track the number of olive trees 13 over large areas. 14

. Location of the study area in Italy (on data from gadm.org). The dark blue area indicates the area in the Italian region of Apuglia where Xylella fastidiosa is present. The dot indicates the study area.
The extent of the damage caused by Xylella fastidiosa is currently poorly understood as the 28 anticipated impacts on regional olive oil production have not yet been reported, thus preventing an 29 assessment of the economic impact. Such assessment could be expedited if the number of trees lost to 30 the epidemic (either directly or indirectly, through preventive phytosanitary measures) were monitored. 31 However, this number is currently unknown. Moreover, information on the number and location of 32 olive trees lost as a result of the Xylella fastidiosa outbreak is required to understand the spread of the 33 pest thus far, and to inform epidemiological models that may simulate it [5]. Knowing the impact of 34 the epidemic on olive tree stocks can also support planning of the response to the emergency, e.g. by 35 informing a more precise evaluation of the impact of the possible containment measures. Finally, an 36 inventory of olive trees could be used to estimate and compensate losses to the olive sector. Given 37 the abundance of olive trees in the Apuglia region, and the high number of landowners, automated 38 analysis of remotely sensed images could generate such inventory most efficiently. 39 Remote sensing based methods for tree counting generally combine image analysis methods with 40 either probabilistics or some form of mathematical morphology (for a review, see [6]) . Most of these 41 studies analyze Red-Green-Blue-NIR or other multispectral data from very high resolution satellite 42 data, but see [7] for an example of the use of both radar and passive optical data for counting trees of 43 various species, including olives. In contrasts, studies for wall-to-wall detection of individual trees in 44 aerial photographs are rare, despite the abundance of such data owing to their use in photogrammetry.
trees have a very constant size, similar shape, and follow a reticular layout. These properties translate 66 into the identifiable co-linearity patterns and patch sizes that have been used thus far to identify olive 67 orchards in satellite images.

68
The resolution of satellite data can be insufficient to reliably discriminate individual olive trees 69 when they follow an irregular grid, particularly when they vary greatly in size and shape, as occurs  Here, we build on image analysis techniques from different scientific fields to develop a method 77 to automatically detect and count individual olive tree crowns in aerial images. We apply the method 78 in an area where olive trees are planted irregularly and are of highly variable sizes, including some of 79 which are old and have very heterogenous crowns. The errors in the automatically obtained counts are 80 then quantified using photo-interpreted reference data. To further assess the robustness of the new 81 method, the validation is replicated using images acquired by different sensors with their own spectral 82 specifications, and taken on different dates and at different spatial resolutions.  The individual samples measure 80 m across and illustrate the heterogeneity of the orchards, including, near to each other, trees with large crowns that form a semi-closed canopy with adjacent trees, trees with crowns that are split into two or more clusters of branches, trees of varying sizes, and trees that have been recently pruned. In addition, they illustrate differences in the understorey and shadowing depending on phenology and image acquisitions times, as well as effects of image resolution. 84 We analyzed four RGB-infrared aerial images: 1) A 50 cm resolution image acquired as part of

92
The camera is gyro-stabilised via a mount with adaptive control. The ADS40 and ADS80 are precursors 93 of the ADS100. The ADS40 relies on a 12-bit Dynamic range CCD and produces 8-bit compressed 94 output. The CCD has a total of 12 lines, four panchromatic, and 8 multispectral, of 12,000 pixels each.

95
The ADS80 improves on the ADS40 by adding double spectra lines for each of the visible and the NIR 96 bands, thereby increases radiometer accuracy. The sensor of the DMC III camera has 25,728 pixels 97 across track and 14,592 pixels along track. It has four colour channels (R, G, B, NIR) recording in 14-bit.

98
It was flown at an altitude of 5000 m.a.s.l. to generate images with a 20 cm pixel size. (1:10 000) map of olive orchards obtained from the regional authorities through its online GIS portal   Next, we removed segments that were smaller than a predefined area corresponding to the 127 smallest crowns (< 0.64 m 2 ), and identified those that were too big to represent a single crown (> 18 128 m 2 , see section Identifying the centres whose crowns touch each other). The remaining segments were 129 then assigned individual IDs, so they could be counted (Figure 7). These operations were carried out

185
Our automated algorithm counted between 32,600 and 42,394 olive trees in the ca 400 ha zones   Total olive tree counts for the study area indicate a greater number of trees in Zone 1 than in Zone

232
Estimates of tree numbers over large areas usually rely on combining estimates of tree cover 233 density and tree cover extent [28]. Here, instead, we present a method to directly, rather than indirectly,

247
Intensely managed olive orchards tend to be even-aged, resulting in trees with similar crown 248 sizes that are spaced along a relatively regular grid. In our study area in Apuglia, orchards are not intensively managed and some trees are more than 100 years old, while others were recently planted. Local heterogeneity in tree sizes also exacerbates the difficulty of counting trees whose crowns 276 form part of a semi-closed canopy, as no a-priori crown size can be assumed for individual trees.

277
Our algorithm singles out image segments larger than 18 m 2 and analyzes whether they are made 278 up of crowns of multiple trees. When this does not appear the case, the segment is counted as a 279 single crown. Some previous algorithms, and particularly OLICOUNT [8] were parcel-driven and 280 semi-automatic, allowing an operator to set estimates of typical crown sizes for individual parcels.

281
While manual tuning of the algorithm on a parcel-basis is unfeasible for wall-to-wall mapping, future 282 development might investigate automated estimations of local crown size distributions to cope with 283 the heterogeneity of olive tree sizes in Apuglia.

285
Here we have shown that it is feasible to count olive trees a landscape characterized by olive 286 orchards with high heterogeneity, both between and within orchards, using aerial photographs.

287
The method developed relies on a series of image segmentation and filtering operations, followed 288 by iterative morphological analyses. Our results indicate that the method, while requiring further 289 development to deal with particular plantation types, can be used to document both spatial and 290 temporal variability in tree crown abundance. Indeed, the method is demonstrated using aerial images 291 of different spatial resolutions, sensor characteristics, and seasons.