Comparative assessment of Deep Learning and Random Forest methods for urban land cover classification (A case study Tabriz city)

Rapid urban growth, especially in developing countries, is causing a large number of urban planning problems. Although only three percent of the global land surface is covered by urban areas, approximately 54% of the world’s population lives in urban centers; according the latest estimates, by 2050 it will increase to nearly 65%. Accurate information on Urban Land Cover (ULC) types and their spatial distribution are of paramount importance for urban planning and management. To date, many studies have been conducted in the context of ULC mapping, and several methodologies and datasets have been used (e.g. land surveying and satellite data) in this regard. Under this background, generating ULC maps using land surveying method is considered as the most accurate technique, however, it is a costly and time-consuming task. Spending the least time and cost to produce these maps is one of the main challenges for city managers. To address this issue, the integration of satellite images and state-of-art classification methods has been received considerable attention in recent years. This study seeks to produce a 10 m resolution ULC map for Tabriz city, locating North East of Iran, using Sentinel-2 satellite data. The present study also aims to compare the potential of two advanced classifiers including Random Forest (RF) and Deep Neural Network (DNN) in ULC mapping. Five ULC classes including bare land, built-up areas, road, vegetation, and water were considered in this regard. As the number of trees (ntree) and the number of variables (mtry) are two main criteria applying the RF algorithm. In this study, ntree was set to 100 and the mtry was set to the square root of the total number of input features. In the case of DNN, a DNN model with six layers, including one input layer with 10 neurons (bands 2-8A and 11-13 of sentinel-2), four hidden layers with 200 neurons per layer, and one output layer (five ULC classes). In this study, the ReLU activation function was used for the hidden layers, softmax activation function was used for classifying information in the output layer. Our findings illustrated that the DNN algorithm by providing 95.2% overall accuracy outperformed RF (overall accuracy = 93.1%). Analyzing the performance of two algorithms regarding ULC classes showed that the DNN algorithm provided better results in bare land and built-up classes; the user’s accuracy and producer’s accuracy of bare land class were respectively 9.6% and 1% higher than those of RF. Regarding the built-up class, these metrics were also higher than RF (user’s accuracy = + 0.3% and producer’s accuracy = + 4.3%). In contrast, the RF algorithm performed better in extracting the road class; the user’s accuracy and producer’s accuracy of road class were 3.65% and 4.1% more than those of DNN, respectively. RF and DNN showed the same performances in classifying vegetation and water classes. In general, both algorithms provided good performances in ULC classification, however, the overall performance obtained by the DNN algorithm was substantially higher than RF. Because the performance of the DNN algorithm is better than the RF algorithm, we concluded that DNN is a valid alternative tool that should be considered for ULC mapping.