Extraction and utilization of visual features for determining plankton concentration levels from aerial images of lake water surfaces

• Main Authors: Desingco, John Darill B., Ngo, Aaron Spencer K.
• Format: text
• Language: English
• Published: Animo Repository 2015
• Subjects: Algal blooms—Monitoring > Remote sensing; Remote sensing
• Online Access: https://animorepository.dlsu.edu.ph/etd_bachelors/12157
• Institution: De La Salle University

Harmful algal blooms (HABs) cause multiple problems all around the world. Algal blooms are known to cause toxicity in the water, fish kills and biomass destruction. This leads to losses in the fishing industry, tourism and the natural environment. Early detection and monitoring of HAB behavior is necessary to assess its impact on its surroundings. In recent years, remote sensing using satellite imagery and complex sensors has become popular for algae monitoring. Implementation of a smaller scale remote sensing system allows observation of less prominent bodies of water and allow flexibility in terms of the time and duration of data collection. The system is designed to mimic the functionality of satellite remote sensing systems for algae detection but using only a normal camera. The system is intended to model the correlation of concentration levels of algae with the visual features of water surfaces using regression analysis. However, visual color analysis does introduce other variables that may interfere with the visual analysis such as uneven illumination and varying lighting conditions. A methodology is proposed to determine the concentration levels by correcting for the weaknesses introduced by visual color analysis and creating a predictive model for future reference. Due to the scarcity of data, stemming from unexpected circumstances, data had to be simulated using floral foam particles as a substance with similar properties as algae particles. The system shows an average of 20% error in prediction for the simulated data. The results indicate that the information in a single and most significant color component was found to be insufficient for determining concentration levels at higher values. This can be seen as either the limitation of single variable analysis on color information for this application or that the samples reached maximum saturation since the value at which each model's accuracy drops is consistent. Color correction of the image sets produced models with increased consistency and increased performance at estimating higher concentration level values. Though the color correction used did not completely remove the effects of lighting from each image, it did reduce the differences between the images under different lighting conditions. A different model for differing light conditions is recommended. Results show that prediction is less consistent on images with uneven illumination and shadows. Data collection in bright and evenly lit environments is advised.