luchayward / csc2005z-treeclassification

Second year research project: Tree presence classification using ML models

CSC2005Z-TreeClassification

Second year research project (CSC2005Z at UCT): Tree presence classification using ML models This project explores the effectiveness of Support Vector machine and Random Forest classifiers for tree detection at the per pixel level using python.

Working with drone scan data provided by Aerobotics, a comparison was performed of the value to use SVM and Random Forest classifiers to create binary masks of tree locations in a scan using RGB and NDVI data.

Abstract:

The abundance of relatively low-cost consumer drones on the market has opened up new opportunities for high quality aerial scans to be captured. Analysing these scans by hand is both time-consuming and error prone. This report examines the viability of using machine learning models to generate bitmasks representing the location of trees in an image using a combination of colour photography and near infrared data. Support Vector Machines and Random Forest classifiers were chosen as they are popular for use in binary classification scenarios such as this. A combination of analytical tests and qualitative comparisons against manually generated ground truths were used to evaluate the effectiveness of the classifiers. It was found that the Random Forest classifier produced the most accurate results without compromising on training time for the model.

Conclusion:

Initial visual comparisons suggested that both models were able to outperform the typical standard that had been deemed acceptable for the hand created masks. Further experimental tests showed that the initial visual perceptions were correct and that both SVM and Random Forest classifiers produced excellent results. By comparing the results of both classifiers, it is clear that in the current use case, both models produce masks which are quantitatively similar to within a margin of error. However, the Random Forest model trained faster and produced high quality results in significantly shorter amounts of time. This makes it a far better solution as the rate of data collection and the resolution of that data is only likely to increase in the future.