Sometimes corporate mottos and slogans, such as our own “Transforming the way the world works”, do not seem to relate to our daily grind, but after reading the publication from Gonçalves et al. 2020, I walked away with deeper understanding of the contributions I make, however small they may be.
I was raised knowing that littering is a tabu and often participated in local Green Up Day activities in the state of Vermont. Unfortunately, this mindset is not as popular in other parts of the world and our oceans and waterways have become dumping grounds transporting the litter to distant shores.
The author’s paper, entitled Quantifying Marine Macro Litter Abundance on a Sandy Beach Using Unmanned Aerial Systems and Object-Oriented Machine Learning Methods, was published in the MDPI remote sensing journal and explores the use of a UAS-based approach for automatically detecting and classifying marine litter.
The Cabedelo Beach was selected as a study area. It is a sandy North Atlantic Ocean beach located on the Portuguese coast. The authors defined a training area and two validation areas for their research.
The RGB aerial imagery was acquired with a Phantom 4 Pro and the final image resolution was 5.5 mm. In addition to the optical data, a DSM was derived from a 3D dense point cloud.
The classification of the data considered 4 classes, defined as follows by the authors:
- Marine Macro Litter (MML) – Persistent, manufactured, or processed solid material
- Vegetation debris – Non-anthropogenic (vegetation) debris
- Dry sand – All kind of dry sand located in the back shore
- Cast shadows – Shadows of all kind of elevated objects and footprints
In order to classify the 4 classes, the authors decided to implement a machine learning approach within eCognition using the Supervised classification algorithm. Supervised classification techniques require training samples as well as “careful selection of discriminating features” so the model(s) can accurately learn the difference between the target classes. In addition to the limiting RGB data, the authors also chose to include additional color spaces: hue-based (HSV = hue saturation & value), perceptually uniform (CIE-Lab) and luminance-based YCbCr.
This resulted in 12 image features that would be used in the upcoming OBIA classification:
- Bands 1-3 = RGB
- Bands 4-6 = perceptually uniform (CIE-Lab)
- Bands 7-9 = hue-based (HSV)
- Bands 10-12 = luminance-based YCbCr
The classification was based on image objects generated with the Multiresolution segmentation algorithm. Gonçalves et al. chose to investigate 3 supervised classification models during the study: 1) Random Forest (RF), 2) Support Vector Machine (SVM) and 3) K-Nearest Neighbor (KNN). eCognition includes 5 different supervised classification models, beyond the 3 looked at by the authors, users can also utilize Decision Trees and Bayes.
As each model has different parameters a degree of “tuning” has to be considered when applying. The authors started with the default parameters and then modified the primary parameters one by one to determine the best results. To assess the performance of their models, the results of the classification(s) were compared to validation data “screened and manually processed by an operator in the GIS environment” – the centroid of each object automatically classified as MML was compared to the centroids of the manual groundtruth objects and if the distance between the two was less than 20 cm the object was deemed correctly classified.
Despite the visually similar classification results, a detailed analysis of the different modes revealed that the RF results had the highest number of True Positives (TP), whereas the SVM model resulted in the lowest number of False Positives (FP) and accordingly the highest precision (77%). The overall F-score (a measure of classification accuracy) for the 3 models were as follows: 65% for KNN, 68% for SVM and 72% for RF.
Along with the identification of MML, the quantification is also extremely important. In this study kernel density estimators (KDEs) were used, whereby the object centroid is transferred with a KDE function to a continuous surface that represents the number of MML items per square meter. The results of the supervised classification also demonstrated a “strong correlation” with the manually generated abundance maps with r-square values of 0.79 (RMSE 0.028 items/m2 ) for RF, 0.76 (RMSE 0.027 items/m2 ) for SVM, and 0.83 (RMSE 0.026 items/m2 ) for KNN.
All in all, the authors note that the RF classifier “obtained better results”, but the KNN model resulted in the “best correlation factor”. The transferability of the supervised classification method also proved positive and the authors feel their approach can be easily rolled out a a production level given the batch processing benefits of eCognition Server.
The OBIA classification approach based on the proposed nomenclature was efficient for extracting the MML class and proved to be well suited for transferring the process to other orthomosaic areas.
It is great to see the success of this study as we are seeing growing interest in the detection of sea/coastal litter using remote sensing tools. I look forward to hearing more from these researchers in the future!