Created by W.Langdon from gp-bibliography.bib Revision:1.7325
To optimize Parthenium mitigation, it is necessary to accurately monitor its spread by using cost-effective solutions, such as remote sensing technologies. In this regard, Sentinel-2 and Landsat 8 imagery, which are freely available, were implemented to accurately map Parthenium infestations. However, several challenges related to the mapping of landscapes infested by Parthenium weed, using conventional classifiers, in combination with Sentinel-2 and Landsat8 imagery, have been overlooked in past studies. For instance, the application of a classifier, which is independent of data characteristics, has not been explored. Meanwhile, it is still not known, which dimension reduction algorithm is appropriate for discarding redundant features from the large volume of Sentinel-2 image data that can be acquired or derived in mapping Parthenium weed. Furthermore, the determination of the temporal window(s) within which the variability of phenological characteristics of Parthenium weed and associated species is the most prominent, and subsequently from which, most accurate maps can be derived, has been overlooked. Therefore, this study endevoured to tackle these issues in order to optimize a Sentinel-2 and Landsat 8 image for more accurate spatial detection of Parthenium weed.
In the first part of this study, the potential of an automated machine learning approach, the Tree-based Pipeline Optimization Tool (TPOT),was explored in mapping Parthenium weed infestations. It was established that the TPOT is an efficient method for automatically selecting and tuning algorithms for Parthenium weed discrimination and monitoring, regardless the data iii characteristics. The TPOT model yielded higher overall classification accuracies(88.15 percent and 74 percent) than the most robust classifier after manual optimization (84.45percent and 68.3percent), using a Sentinel-2 and Landsat 8 images, respectively.
Secondly, ten feature selection algorithms, which belong to five groups, namely, sparse learning-based, statistical-based, information theoretical-based, similarity-based and wrappers methods, were compared on Sentinel-2 wavebands and their derived vegetation indices in mapping Parthenium weed, using specific class-based accuracy metrics. The results showed that the investigated feature selection algorithms could increase the classification accuracies of Parthenium weed, in addition to reducing the number of variables or features. The svm-b, a wrapper method, produced the highest classification accuracies, and ReliefF, a similarity-based method, could select the smallest size of the optimal features.
The third part of the study endeavoured to find the temporal window within which variability in the phenological characteristics of Parthenium weed and its associated species is the most asynchronous, and subsequently an accurate map of Parthenium weed can be derived using a Sentinel-2 image. The results showed that most accurate maps of Parthenium weed could be obtained at the beginning of February. Bands such as Blue (490 nanometres), NIR (835 nm), Red-edge (704 nm) and Green (560 nm) were the most contributing features in the developed models.
In the fourth part of the study, a hybrid feature algorithm was proposed for handling the correlated variables in a multi-date Sentinel-2 image. The proposed approach, which combines ReliefF, svm-b and RF, was compared against its constituent feature selection methods. The multi-date and the single-date images acquired at the beginning of February were also compared. The results showed that the proposed feature selection algorithm selects fewer features than the single feature selection methods, in addition to producing higher classification accuracies (e.g. Overall Accuracy, Producer and User Accuracies) than the single-date image. The Overall Accuracy was 86.6percent, with 22 optimal features using the proposed approach, whereas it was 84.7percent with 35 optimal features using svm-b, 84percent with 31 optimal features using ReliefF, 85percent with 38 optimal features using RF and 77.6percent using the single-date image.
Finally, a hybrid feature selection algorithm and the TPOT were combined in a new algorithm system to explore the capability of the TPOT for handling high dimensional geo-datasets, such as the multi-date Sentinel-2 image. The results showed that the TPOT can be applied on high dimensional datasets without affecting the classification accuracies. The highest Producers and Users accuracies of Parthenium weed were achieved, using a multi-date image in combination with the TPOT (90percent and 93percent). Coupling feature selection with the TPOT reduces the computational costs (17percent) at the expense of the classification accuracies.
Overall, this study has proved that, by overcoming some previously overlooked challenges related to weed mapping, a Sentinel-2 image can be optimized and hence, significant improvement of the spatial representation of Parthenium weed in infested landscapes can be achieved. Information on the accurate extent of Parthenium weed is crucial for enhancing decision-making in the management plans.",
Genetic Programming entries for Zolo Zime Zinu Serge Kiala