An Automated Open-Channel Encroachment Identification Model Based on Machine Learning in Los Angeles County

Introduction: The open channels of Los Angeles County are prone to encroachments such as vegetation, encampments, and other objects that impede water flow. Routine inspections are required to identify and address encroachments. However, the traditional encroachment inspection approach includes an on-site survey conducted by engineers, which is both time intensive and incurs high labor costs. Figure 1 shows an example of a vegetation encroachment. Therefore, a new encroachment identification approach is developed in this paper, based on machine learning and remote sensing technologies. Remote sensing technology collects aerial images of reflectance measurements from different bands of the electromagnetic spectrum. Reflectance from target land cover classifications (e.g., vegetation encroachment, open channel, water, etc.) can be used as training data for a machine learning model. Once the machine learning model is trained, it can then be applied to reflectance measurements across a given area of interest to identify encroachment on a pixel-by-pixel basis. Most machine learning algorithms require robust training data datasets in order to reach satisfactory model performance. In this paper, 10 open channel section aerial photos were collected and labeled by the Stantec project team, resulting in about 20 million total pixels to be used as training points. The model is trained on those 10 open channel section aerial photos, with each photo containing four bands (red, green, blue, and near infrared). The model prediction performance is evaluated by accuracy, F1-score, and encroachment class recall values. The preliminary results indicate that the model can reach an overall 74% prediction accuracy with 0.51 F1-score. Around 73% of the vegetation encroachments in the original photos were correctly captured by model predictions. Methodology: The data used for training the machine learning model includes 10 open-channel section aerial images in Tag Image File Format (TIFF). TIFF files were acquired from the Los Angeles Region Imagery Acquisition Consortium (LARIAC) Program. TIFF images store pixel values (e.g., different reflectance values for each wavelength band measured) and geo-reference information. The pixel size is 0.25 by 0.25 ft. Figure 2 shows an example of the training data, where Figure 2a) shows the original aerial image and Figure 2b) shows the pixels designated as Clean Dry Open Channel pixels. The original aerial data need to be labeled for downstream analysis. Imagery pixels were labeled as one of five potential classifications. The five classifications included: 1) Clean Dry Channel, 2) Clean Wet Channel, 3) Others, 4) Overhead Structure, and 5) Vegetation, as shown in Table 1. Labels were created by visually identifying the five different classifications from imagery and manually drawing polygons around the different classes with a GIS software. Figure 3 is an example showing partial polygons drawn for an aerial image. The TIFF images used in this study contained the reflectance values from four bands of the electromagnetic spectrum: red, green, blue, and near infrared. Band values range from 0 to 255, representing the brightness value for each band. Figure 4 provides an illustrative example of different bands. The four individual reflectance bands will serve as four features in the machine learning model. In addition to using reflectance from each individual band, we also calculated indices from these bands and included them as features in the model as well. The creation of new features, also known as 'feature engineering', is often required in machine learning. Many indices are used in remote sensing to help identify different land cover types based on the unique reflectance patterns they have. In this project, we will use four remote sensing indices: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), and Normalized Difference Water Index (NDWI). Classification and Regression Trees (CART) is selected as the machine learning model in this project because it is able to explain how a target variable's value (label) can be predicted based on other values (features). Model Results: One training image is chosen to evaluate the model performance on training data. Figure 5 shows the comparison between the true label (a) and the predicted label from the model (b). For all five classes, the predictions closely align with the actual labels. The Vegetation and Overhead structure pixel crowds are well captured. The training metrics are calculated as 81% accuracy, 0.55 F1 score, 0.86 Vegetation recall, and 0.92 Others recall. Although the training metrics cannot be the sole indicators of model performance, they can still illustrate how well the model performs and demonstrates its potential to be powerful on the test dataset. The model is also evaluated on a single test image. The confusion matrix result is shown in Figure 6. The test metrics are calculated as 74% accuracy, 0.51 F1-score, 0.73 vegetation recall, and 0.18 Others recall. In this section, all the test data come from the same TIFF image, we are able to visualize and compare the predictions vs. true labels. Figure 7 shows the test image labeled using the true labels, predicted labels, real TIFF data, and real vegetation class polygons in subfigures a), b), c) and d) respectively. Comparing subfigures a) and b), it is shown that class 1 (Clean Dry Channel) and class 2 (Clean Wet channel) are predicted well, while class 4 (Overhead structure) is partially captured. Since class 5 (Vegetation) is hard to see in the true label plot due to the overlap with class 2 (Clean Wet channel) label, subfigures c) and d) are included. Subfigure d) is based on the true TIFF image c), where polygons of vegetation class are drawn and visible. Most of the true vegetation pixels are in the middle of the channel, which is reflected in the prediction figure subfigure b). The model fails to capture the Others class (class 3) very well, as illustrated in the low recall value. In fact, a high Other recall on training data (0.92) and a low Other recall on test data (0.18) indicates that the model has potential but requires more information on the Others class (such as the type of encroachment). However, the vegetation encroachment can be predicted fairly well. Conclusions: The objective of this paper is to develop a machine learning model to automatically identify vegetation encroachments in open channels in Los Angeles County. The machine learning model developed is based on the CART model, which is a simple white box model that can handle complex data. The model is trained on 10 open channel section aerial images, where each image contains four bands (red, green, blue, and near infrared) in the TIFF file format and has been manually labeled using polygons to categorize each pixel into an encroachment class. The model prediction performance is evaluated using accuracy, F1-score, and encroachment class recall values. The preliminary results indicate that the model can reach an overall 74% prediction accuracy with a 0.51 F1-score. In addition, the Vegetation class recall score means that 73% of the vegetation encroachments in the original photos can be correctly captured by model predictions.