Title
Cargo Measurement in the Port of Oakland using Classification

Author Information
James Beggs
American River College, Geography 350: Data Acquisition in GIS; Fall 2018
w1495888@apps.losrios.edu

Abstract
The goal of this study was to measure the volume of cargo that is passing through the Port of Oakland. This paper explores the feasibility of using classification tools and zonal statistics measure shipping containers and cargo stored in the international section of the Port and measure the square footage of cargo passing through the port. A key concept for this project was the ability to show a relationship between asphalt coverage and cargo coverage. Heavy focus was placed on developing a process that could be replicated. The project was split into a few different steps developing

Land use Schema

Refining the correct spectral, spatial resolution, and pixel size

Unsupervised classification

Zonal Statistics

Results illustrated that using a standard classification system in ARCpro is not ideal for cargo segmentation, and classification but the system did provide the expected results. Results showed an increase of cargo between July and November, as expected and further refinement of the process is needed to accurately determine the square footage of cargo in the Port of Oakland. Images were obtained from Planet and is 3 meter by 3 meter resolution. Boundary for the Port of Oakland was user generated for this project and referenced existing port maps to refine the boundary for the international section.

Introduction
With current development daily imagery that is not available to people increased interest has been shown from financial institution and real estate firms to use this data to help determine the financial feasibility of a property or company. Current trends use people to analyze the data, company "black Box" with unknown process, or machine learning. Each of these processes are expensive and opaque when compared to a more simple process like land use classification. The Port of Oakland was selected for a few reasons. First daily imagery is available free from Plant for this location, Second the actual number of the tonnage of cargo passing thru the Port is available for us to check the accuracy of our numbers.

Background
Two images were selected one in July a section of the year that has lower consumer levels, and then another in November which would be expected to have higher levels of consumer behavior as Christmas season buying is taking place.

Methods
The tools used were the imagery processing tools in ARCpro 2.2.1. These tools allow for Raster clipping, segmentation, land use schema development, unsupervised classification, and zonal statistics.

Analysis
The first part was identifying the port boundaries and clipping the raster. This process allows just the port and the cargo to be focused on. Reducing the chance that random building, cars, or other features of the environment would confuse the segmentation or classification process reducing the ability to get an accurate measurement of cargo and accurate ratios of asphalt to cargo.


The relationship between asphalt and cargo is expected because as cargo is brought in it is stored on top of the asphalt and reduces asphalt coverage and increases cargo coverage. Next step required establishing a segmentation setting that would detect the difference between cargo and asphalt. Segmentation process has three specific settings spectral detail, spatial detail and the minimum pixel size. Being that cargo is multicolored and can be many different shapes the segmentation process focused on getting as much asphalt as possible into a single segment as possible. This was to reduce the possibility of having asphalt classified as cargo and skewing cargo measurement upwards. During the segmentation process settings for minimum pixel size were tested between 20 and 1. The best setting was minimum segment pixel size was 4 pixels, anything lower created too many segments to not be useful. Below is an example of minimum pixel size set to 1.



Spectral detail was set to be higher at 14, and a Spatial detail was set to 18. When using these settings on the July imagery they did the best job at breaking asphalt into a single segment when compared to segmented July imagery with different settings on pixel size, spectral detail and spatial detail. Below is an example of a segmented process with the at pixels 5, Spatial Detail at 19 and Spectral Detail 17.



Below is the selected segmentation settings at Pixels at 4, Spatial Detail at 18 Spectral Detail at 14. The main advantage of these settings was that it captures most of the asphalt into a single category.



The next step in the process was to classify the images. This process presented problems, because up to this point all steps had been done manually. The classification process does not lump similar segments together as expected but can result in unexpected segments being classed different than similar segments that are next to them as illustrated below where asphalt and water has been classified incorrectly even thought they were segmented properly.



This could be an issue with the maximum likelihood process that was used, these settings were not studies and only maximum likelihood was used. Each step in the process reduces the image detail and once classification image detail could be too abstract to accurately represent the amount of cargo that was on site. Classification training in ARCpro is done through the development of polygons that are saved into a shape file. These polygons are then used to develop "samples" for the class. This process is not ideal with quickly changing landscapes. To resolve this issue the process of an unsupervised classification was used. This process did not require a shape file to select training samples, all training samples were selected by the software. Unsupervised classification worked well because asphalt will always be the largest land category and is in predictable locations so a user can determine which class is asphalt and which is cargo.



The final step of this process is calculating the Zonal Statistics. After reaching this step I then had a workflow developed and used this process on the November Imagery.

Clip raster
Segment image
Run unsupervised imagery classification
Zonal Statistics





Results
Results looked marginal, it did miss classify a large section of cargo as asphalt. As a final step I generated statistics for each classified image. The result of the process is that additional understanding of this tool is required. Pixel count form the process July imagery to the November imagery was not the same even though the same pixel size, satellite, and boundary was used.
July Imagery Class Count Class Type Sum of Count
0 111116 Asphalt 111116
1 14752 Water 14752
2 6665 Cargo Blue 17861
3 11196 Cargo White See Above
Count Sum 143729



November Imagery Class Count Class Type Sum of Count
0 5156 Cargo 1 See Below
1 12907 Cargo 2 1See Below
2 1000861 Asphalt 1000861
3 7596 Cargo 3 25659
Count Sum 1043729

Differance Between July and November 7798

Conclusion
With this process I was able to illustrate that there was an increase with the number of pixels classified as "cargo" comparing July imagery to November imagery. The end result of this process is that further refinement of the process is required. Although there was a detected increase from the July imagery to the November imagery, and there was a workflow developed that could be automated with the workflow builder, the process is not accurate enough to be able to develop an accurate measurement of the true amount of cargo passing thru the port.