Evaluation of the Impact of Image Spatial Resolution in Designing a Context-Based Fully Convolution Neural Networks for Flood Mapping

Abstract

In this paper, our main aim is to investigate the context-based pixel-wise classification using a fully convolutional neural networks model for flood extent mapping from multispectral remote sensing images. Our approach helps to overcome the limitation of the conventional classification methods with low generalisation ability that used per-pixel spectral information for pixel-wise classification. In this study, a comparative analysis with conventional pixel-wise SVM classifier shows that our proposed model has higher generalisation ability for flooded area detection. By using remote sensing images with different spatial resolutions we also aim to investigate the relationship between image-sensor resolution and neighbourhood window size for context-based classification. Instead of fine-tuning a pre-established deep neural network model, we developed a preliminary base model with two convolutional layers. The model was tested on images with two different spatial resolutions of 3 meters (PlanetScope image) and 30 meters (Landsat-5 Thematic Mapper). During training phase we determined the structure of the convolutional layer as well as the appropriate size of the contextual neighbourhood for those two data types. Preliminary results showed that with increasing the scale of spatial resolutions the required neighbourhood size for training samples also increases. We tested different neighbourhood sized training samples to train the model and the analysis of the performance of those models showed that a 11 × 11 neighbourhood window for PlanetScope data and a 3 × 3 neighbourhood window for Landsat data were found to be the optimum size for classification. Insights from this work may be used to design efficient classifiers in scenarios where data with different resolutions are available.