Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Titre : Encoding remotely sensed time series data as two-dimensional images for urban change detection using convolution neural networks

Auteur : Dukes, Marc Daniel

Université de soutenance : Stellenbosch University

Grade : Master of Science (MS) 2022

Résumé partiel
Urban expansion is the most pervasive form of land cover change in South Africa. A method that can effectively detect and indicate areas that have a higher probability of displaying urban change will therefore be a valuable asset to analysts. That is why it is critical to derive a rapid framework that can accurately map urban change. An alternative remote sensing approach that uses multi-temporal time series data and deep learning techniques has been proposed as a potential method for performing a successful urban change detection. The interdisciplinary scientific field of computer vision holds a framework for encoding time-series data as two-dimensional (2D) images for input to a convolution neural network (CNN). Traditional image classifications techniques and more recent studies that have deployed machine learning and deep learning classifiers (namely support vector machine (SVM), random forest (RF), k-nearest neighbour (kNN), long short-term memory (LSTM) and CNN) have been used for urban land cover classification. In this study, a unique framework proposed within computer vision that exploits Gramian angular fields (GAF) and Markov transition fields (MTF) as the transformations for encoding time series data as 2D imagery prior to deep learning classification is investigated for urban change detection. Two main experiments were carried out, both of which utilised the proposed framework for performing an effective urban change detection. The first experiment used coarse resolution data derived from Pretoria using MODIS 500m and 250m normalised difference vegetation index (NDVI). The proposed framework was then deployed, and Gramian angular summation field (GASF), Gramian angular difference field (GADF), and MTF transformations used to encode the time series data. A concatenated encoded image containing the information from all three transformations was formed and was run alongside the three individual transformations. Multiple pre-trained CNN architectures (namely ResNet, DenseNet, InceptionV3, InceptionResNetV2, VGG and MobileNet) were used, from which an urban change detection was derived. It was established that the concatenated images yielded the highest accuracy at 91% and 93% for the 500m and 250m resolution datasets, respectively. The proposed framework was compared to a current state-of-the-art time series classifier (LSTM) to illustrate the effectiveness of encoding and processing deep learning classifiers. The results also outperformed that of other urban change detections studies conducted in South Africa. The second experiment made use of higher resolution Sentinel-2 data derived from a resampled 30m resolution NDVI product of Pretoria. Several investigations were made into the influencing elements that affect the performance of the urban change detection.

Présentation

Version intégrale (3,22 Mb)