Massey Documents by Type

Permanent URI for this communityhttps://mro.massey.ac.nz/handle/10179/294

Browse

Subject: search.filters.subject.460303 Computational imaging

Search Results

Now showing 1 - 5 of 5
  • Thumbnail Image
    Item
    Landscape configuration and composition trends of indigenous forest using multi- and hyperspectral imagery : a thesis presented in partial fulfilment of the requirements for the degree of Master of Science in Geography at Massey University, Manawatu, New Zealand
    (Massey University, 2023) Carter, Stefan Rex
    Historical land use in New Zealand led to the widespread destruction of indigenous forests, creating a complex landscape matrix within which forest fragments now survive. Legislative changes have effectively halted large-scale deforestation, yet small-scale fragmentation continues to occur. This negatively effects numerous ecological processes vital in maintaining healthy forest structure and function. To monitor fragmentation, state of the environment reports often quantify area and percentage of landscape change. While these metrics are important, they are poorly suited to quantifying the spatial configuration and distribution of forest fragments. To quantify the spatial arrangement of indigenous forest remnants, two case studies were undertaken at different spatial and temporal scales. The first employed low-resolution (246 m²), multispectral derived imagery from the Climate Change Initiative - Land Cover (CCI-LC) dataset to quantify regional fragmentation over a 28-year period. Landscape analysis was performed with FRAGSTATS and revealed both spatial and temporal variability in the arrangement of forest patches. This was evidenced by periods of fragmented growth and decline, along with periods of infilling in some regions. Despite several regions recording a net increase in forest area, the overall trend was towards greater disaggregation and fragmentation. However, it is essential to exercise caution when interpreting these findings as the coarse resolution of the CCI-LC dataset may not adequately describe fragmentation at the regional level. The second case study employed high-resolution (1 m²), hyperspectral imagery to quantify forest fragmentation on a rural property. Imagery was captured with the AISA FENIX hyperspectral camera and atmospherically corrected using a pixel-wise radiative transfer model. Land cover was classified with a one-dimensional convolutional neural network and landscape configuration was assessed with FRAGSTATS. The results were compared to the medium-resolution (1 ha nominal mapping unit) Landcover Database (LCDB) and the low-resolution CCI-LC. Greater accuracy in both land cover classification and definition were achieved with the hyperspectral imagery. Edge-perimeter and connectivity metrics were also substantially improved. Management strategies seeking to reduce fragmentation should consider the use of high-resolution, hyperspectral imagery in conjunction with landscape metrics to improve classification accuracy and precision.
  • Thumbnail Image
    Item
    Hyperspectral remote sensing for early detection of wild carrot in Carrot (Daucus carota) seed production : a feasibility study : a thesis presented in partial fulfilment of the requirements for the degree of Master of Science in Horticultural Science at Massey University, Manawatū, New Zealand
    (Massey University, 2023) Bhatia, Sunmeet Singh
    Carrot (Daucus carota) seed production is an important part of the NZ vegetable seed industry with exports of $33.4 million NZD in 2020. Most carrot seed production is based in the Canterbury region, but there is a desire by key stakeholders to expand carrot seed production in the Hawke’s Bay region of NZ. However, presence of weed wild carrot (Daucus carota subsp. carota) in the region acts as a significant constraint to carrot seed production. Wild carrot plants can crosspollinate with carrot crop plants, causing genetic contamination in a crop where genetic purity is of critical importance. The current weed management strategy of manual scouting and rouging is resource intensive and ineffective in achieving appropriate control of wild carrot in the region. Airborne hyperspectral remote sensing is a technology that has proven its ability in plant species identification and can do so at a high spatial scale in a short period of time. This makes the technology a promising candidate for a superior alternative weed control method. This project aimed to test the feasibility of the technology to identify wild carrot plants in carrot seed crop fields and nearby areas within the crop’s isolation distance (2000m). This involved creating spectral libraries of dominant plant species/materials, including wild carrot, in the area of interest. The methodology involved conducting a survey and collecting airborne hyperspectral data. Further, ground-based collection of GNSS enabled accurate GPS locations of wild carrot plants in the survey area, acted as training and validation data for subsequent classification analysis. The ground truth data was also used for a pixel composition analysis – which also helped understand the environmental context of wild carrot plants. The data was analysed in an image processing software (ENVI®, v5.6). The analysis involved two levels of classification algorithms. A first order classification – minimum distance classification (MDC) – was used to classify the data into broad land surface cover types. The classification was successful with an overall accuracy of 96%. The second order classification was a soft classification algorithm which employed spectral unmixing – mixture tuned matched-filtering (MTMF). This method allows sub-pixel classification when the target surface is smaller than pixel size, as in this case. MTMF helped create a model which predicted potential locations of wild carrot plants at a threshold level of 5% of pixel area (surface area - 0.05m²) and a producer’s accuracy of 70% (Omission error rate – 30%) for patches above the threshold surface area. These predicted locations were projected on appropriate RGB base layers to create wild carrot weed maps. The biggest limitation was likely the 1m² spatial resolution of the hyperspectral camera employed in the study, which dictated the 5% pixel threshold level. These detection threshold and accuracy levels are lower than in other similar studies, however they are likely acceptable in the current context and can help mitigate wild carrot damage in carrot seed production in the Hawke’s Bay. The study has helped identify areas of future research to further improve the detection threshold and accuracy levels. These include identifying relationships between environmental context related parameters and wild carrot manifestation, acquiring higher spatial resolution data (lower altitude flights, unmanned aerial vehicle (UAV) mounted cameras, deploying of image fusion techniques using separate high spectral (hyperspectral) and spatial resolution (RGB/multispectral) imagery.
  • Thumbnail Image
    Item
    Development of a novel methodology for the measurement of Red19 kiwifruit colour quality : a thesis presented in partial fulfilment of the requirements for the degree of Master of Engineering in Mechatronics, Massey University, New Zealand
    (Massey University, 2021) Lockwood-Geck, Samuel A.
    Consumers love visually appealing fruit. Colour is a key aspect of this appeal and is determined by a multitude of physical and psychophysical phenomena. Compared to its green and gold counterparts, the colour of red kiwifruit is not easily measured using existing techniques due to the spatially varying nature of its flesh colour. Red19 kiwifruit’s flesh colour often presents as a mixture of green, yellow, and red to varying degrees. The current method used to measure kiwifruit redness is by a subjective observer, but this produces a noisy and unreliable dataset which makes quality control difficult. The objective of this thesis is to create a colour measurement system for Red19 kiwifruit that transforms the current qualitative system into a quantitative system that can be utilised by the kiwifruit industry to produce standard and reliable measures of kiwifruit redness. Considering this goal, this thesis aims to explore what constitutes an objective grading scale of Red19 kiwifruit, and how a colour measurement system might be created that is invariant to changes in illuminant and spectral sensitivity (observers) and how this system could be deployed on a typical consumer smartphone. This thesis presents a comprehensive exploration of colour constancy in relation to red kiwifruit. A new scale of redness is proposed, and a small sensory trial was undertaken to validate this scale. A dataset of hyperspectral and raw images was collected and utilised to create a model that estimates the average reflectance of a single red kiwifruit. Another model was then created to regress from the average reflectance of a kiwifruit to a final redness score where fruit were graded based on the outcome of the sensory trial. The new redness scale proposed therein spans a range of 0 to 10,000 and is calculated by taking the difference of a kiwifruit’s average red and green RGB channel in the AdobeRGB colour space. This range was also divided by five to produce alternative class-based scale containing five classes each spanning sections of 2000 units over the new redness scale. This discrete scale is similar to the current qualitative scale used by Zespri to assess the colour of red kiwifruit. Through the application of a convolutional machine learning model, the average reflectance spectra of a kiwifruit can be estimated producing an RMSE of 0.0109 and 0.0096 over the visible and visible NIR spectral ranges. A general regression model is then used to regress from a kiwifruit’s average reflectance spectra to a final redness score and this produces a mean absolute error of 213.82 with a standard deviation of 213.82 which is equivalent to an average error of 2.2% and standard deviation of 2.13% when considering the full range of scores on the redness scale. These models are then combined to produce a final model, named KiwiNet, that produces a correct kiwifruit classification rate of 91.06%. This model has been demonstrated to run on a typical consumer smartphone and can produce a reflectance estimate and redness score for each kiwifruit. This model has been demonstrated to be invariant to five different illuminants and twelve different spectral sensitivities. Future work should look to carry out a larger sensory trial and explicitly corelate the proposed redness scale to the existing one. Likewise, this study highlighted that a kiwifruit image taken at 583nm appears to be used by the model to estimate kiwifruit redness score and work to reconstruct this wavelength from RGB/RAW images could provide a single measure of redness in the future.
  • Thumbnail Image
    Item
    Real-time object detection using IoT devices : detecting rodents and birds for conservation : a thesis presented in partial fulfilment of the requirements for the degree of Master of Science in Computer Sciences at Massey University, Auckland, New Zealand
    (Massey University, 2021) Shim, Kyuwon
    In New Zealand forests rodents currently threatens various species of native birds. Unfortunately, the traditional traps and poisoning are not sufficient to eliminate an increasing number of rodents in our forests. In this work we developed an auxiliary control tool that combines the latest computer technology, such as object recognition algorithms, with Internet of Things(IoT) devices. A drawback of most object recognition algorithms is that it requires a tremendous amount of computation, so these techniques are not necessarily adequate to be applied to IoT devices. This thesis proposes to create a real-time object detection algorithm using IoT devices for two classes of objects, rodents and birds. Two methods were implemented and tested. The first method is a shape-based algorithm using Fourier Descriptors (FD). In order to find the ideal classification algorithm for the FD method, we compared the accuracies of four different classification algorithms: Neural Network, Random Forest, AdaBoost, and Support Vector Machine(SVM). The second method is based on YOLO (You Only Look Once). As YOLO requires a huge amount of images to train well, we created a semi-automated labelling system in our detection algorithm to boost the number of images for training. The labelling system produces the training requirements for both Yolo and FD and promises much faster processing time than manual labelling work. The labelling system succeeded in obtaining the labelled data of 2172 rodents and 3494 birds. Using a desktop machine and the FD method, and the average accuracy rate was 80% when using the Random Forest classifier. Using a desktop machine and the YOLO method, the average accuracy rate was 97%. The frame rates were 19 fps for the FD, 3.7 fps for the YOLO on CPUs and 90.9 fps for YOLO running on a GPU. Using a Raspberry Pi 3 and a video dataset, the FD method resulted in an 83% general accuracy and a frame rate of 6.33 fps, which is 30% more accurate and 21 times faster than YOLO. The experiments showed the feasibility of using the FD method in Raspberry Pis for real-time detection, with performance advantages compared to the YOLO architecture.
  • Thumbnail Image
    Item
    GPU accelerated procedural terrain generation : a thesis presented in partial fulfilment of the requirements for the degree Master of Science in Computer Science at Massey University, Albany, New Zealand
    (Massey University, 2021) Kim, Richard Changwoo
    Virtual terrain is often used as the large scale background of computer graphics scenes. While virtual terrain is essential for representing landscapes, manual reproduction of such large-scale objects from scratch is time-consuming and costly for human artists. Many algorithmic generation methods have been proposed as an alternative solution to manual reproduction. However, those methods are still limited when needing them to be employed in a wide range of applications. Alternatively, simulation of the stream power equation can effectively model landscape evolution at large temporal and spatial scales by simulating the land-forming process. This equation was successfully employed by a previous study in terrain generation. However, the unoptimised pipeline implementation of the method suffers from long computation time on the increased simulation size. Graphics processing units (GPUs) provide significantly higher computational throughput for massively parallel problems over conventional multi-core CPUs. The previous study proposed a general parallel algorithm to compute the simulation pipeline, but is design for any multi-core hardware and does not fully utilise the computing power of GPUs. This study seeks to develop an optimised pipeline of the original stream power equation method for GPUs. Results showed that the new parallel GPU algorithm consistently had higher performance (about 300% for GTX 780 and 900% for RTX 2070 Super) recent octa-core CPU (Intel i7 9700k 4.9 Ghz). It also consistently showed a 300% improvement in performance over the previous parallel algorithm on GPUs. The new algorithm significantly outperformed the fastest parallel algorithm available, while still being able to produce the same terrain result as the original stream power equation method. This advancement in computational performance allows the algorithm method to generate precise geological details of terrain while providing reasonable computation time for the method to be employed in a broader range of applications.