Day 4

Detailed paper information

Back to list

Paper title Deep Learning model for automatic detection and mapping of sandy shoreline
  1. Soumia Bengoufa CNRS / LETG UMR 6554 Speaker
  2. Simona Niculescu CNRS LETG UMR 6554 - Institut Universitaire Européen de la Mer
  3. Guanyao Xie CNRS / LETG UMR 6554 Speaker
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text The shoreline is an important feature for several fields such as erosion rate estimation and coastal hazard assessment. However, its detection and delineation are tedious tasks when using traditional techniques or ground surveys, which are very costly and time consuming. The availability of remotely sensed data that provide synoptic coverage of the coastal zone and recent advances in image processing methods overcome the limits of these traditional techniques. Recent advances in artificial intelligence have led to the development of the Deep Learning (DL) algorithm that have recently emerged as a discipline used in image processing and earth sciences. Several studies have used these approaches for feature extraction via image classification, but no study has explored the potential of a DL method for automatic extraction of a sandy shoreline.
The present study implements a methodology for automatic detection and mapping of the position of the sandy shoreline. Thus, the performance of a supervised classification of multispectral images based on a convolutional neural network (CNN) model is explored. Indeed, a comparative study between several robust machine learning (ML) models, namely SVM and RF, was carried out on the basis of the accuracy of the predictive results in a micro-tidal coast such as the Mediterranean coast.
The CNN model was developed for land cover classification (4 classes), designed, trained and applied in the eCognition software, using Pleiades images. Its architectures were designed to meet our objective, which is the detection of a specific target class (wet sand class) with relatively narrow dimensions. Overall, several experiments with different sample patch sizes [(4 x 4), (8 x 8), (16 x 16) and (32 x 32)] were performed to define the number of convolutional layers. Therefore, the architecture of an input layer of (8 x 8 and 4 spectral bands), with three convolution layers and max-pooling after the first layer, was preferred. The hyper-parameters of the model were empirically tuned by a cross-validation process. The results were validated by calculating the distance between the extracted shoreline and the reference line, which was acquired in situ on the same day as the Pleiades image mission.
Overall, all the models performed quite well with an Overall Accuracy (OA) over of 85%. The SVM algorithm achieved a lowest OA coefficient of around 85.8 %, while RF and CNN have achieved 90% and 91.4% respectively. The performance of the CNN model is superior comparing to that of ML algorithms. It is noted that 76% of the extracted shoreline by the CNN model is located within 0.5m of the reference (in-situ) shoreline against 53% and 42% of the shoreline extracted by the RF and SVM algorithms respectively.