Accurate knowledge of urban forest patterns contributes to well-managed urbanization, but accurate urban tree canopy mapping is still a challenging task because of the complexity of the urban structure. In this paper, a new method that combines double-branch U-NET with multi-temporal satellite images containing phenological information is introduced to accurately map urban tree canopies. Based on the constructed GF-2 image dataset, we developed a double-branch U-NET based on the feature fusion strategy using multi-temporal images to obtain an accuracy improvement with an IOU (intersection over union) of 2.3% and an F1-Score of 1.3% at the pixel level compared to the U-NET using mono-temporal images which performs best in existing studies for urban tree canopy mapping. We also found that the double-branch U-NET based on the feature fusion strategy has better accuracy than the early fusion strategy and decision fusion strategy in processing multi-temporal images for urban tree canopy mapping. We compared the impact of image combinations of different seasons on the urban tree canopy mapping task and found that the combination of summer and autumn images had the highest accuracy in the study area. Our research not only provides a high-precision urban tree canopy mapping method but also provides a direction to improve the accuracy both from the model structure and data potential when using deep learning for urban tree canopy mapping.

Automatically extracting buildings from remote sensing images (RSI) plays important roles in urban planning, population estimation, disaster emergency response, etc. With the development of deep learning technology, convolutional neural networks (CNN) with better performance than traditional methods have been widely used in extracting buildings from remote sensing imagery (RSI). But it still faces some problems. First of all, low-level features extracted by shallow layers and abstract features extracted by deep layers of the artificial neural network could not be fully fused. it makes building extraction is often inaccurate, especially for buildings with complex structures, irregular shapes and small sizes. Secondly, there are so many parameters that need to be trained in a network, which occupies a lot of computing resources and consumes a lot of time in the training process. By analyzing the structure of the CNN, we found that abstract features extracted by deep layers with low geospatial resolution contain more semantic information. These abstract features are conducive to determine the category of pixels while not sensitive to the boundaries of the buildings. We found the stride of the convolution kernel and pooling operation reduced the geospatial resolution of feature maps, so, this paper proposed a simple and effective strategy—reduce the stride of convolution kernel contains in one of the layers and reduced the number of convolutional kernels to alleviate the above two bottlenecks. This strategy was used to deeplabv3+net and the experimental results for both the WHU Building Dataset and Massachusetts Building Dataset. Compared with the original deeplabv3+net the result showed that this strategy has a better performance. In terms of WHU building data set, the Intersection over Union (IoU) increased by 1.4% and F1 score increased by 0.9%; in terms of Massachusetts Building Dataset, IoU increased by 3.31% and F1 score increased by 2.3%.