No abstract

Currently, there exists a growing demand for individual building mapping in regions of rapid urban growth in less-developed countries. Most existing methods can segment buildings but cannot discriminate adjacent buildings. Here, we present a new convolutional neural network architecture (CNN) called U-net-id that performs building instance segmentation. The proposed network is trained with WorldView-3 satellite RGB images (0.3 m) and three different labeled masks. The first is the building mask; the second is the border mask, which is the border of the building segment with 4 pixels added outside and 3 pixels inside; and the third is the inner segment mask, which is the segment of the building diminished by 2 pixels. The architecture consists of three parallel paths, one for each mask, all starting with a U-net model. To accurately capture the overlap between the masks, all activation layers of the U-nets are copied and concatenated on each path and sent to two additional convolutional layers before the output activation layers. The method was tested with a dataset of 7563 manually delineated individual buildings of the city of Joanópolis-SP, Brazil. On this dataset, the semantic segmentation showed an overall accuracy of 97.67% and an F1-Score of 0.937 and the building individual instance segmentation showed good performance with a mean intersection over union (IoU) of 0.582 (median IoU = 0.694).

Mapping urban trees with images at a very high spatial resolution (≤1 m) is a particularly relevant recent challenge due to the need to assess the ecosystem services they provide. However, due to the effort needed to produce these maps from tree censuses or with remote sensing data, few cities in the world have a complete tree cover map. Here, we present the tree cover data at 1-m spatial resolution of the Metropolitan Region of São Paulo, Brazil, the fourth largest urban agglomeration in the world. This dataset, based on 71 orthorectified RGB aerial photographs taken in 2010 at 1-m spatial resolution, was produced using a deep learning method for image segmentation called U-net. The model was trained with 1286 images of size 64 × 64 pixels at 1-m spatial resolution, containing one or more trees or only background, and their labelled masks. The validation was based on 322 images of the same size not used in the training and their labelled masks. The map produced by the U-net algorithm showed an excellent level of accuracy, with an overall accuracy of 96.4% and an F1-score of 0.941 (precision = 0.945 and recall = 0.937). This dataset is a valuable input for the estimation of urban forest ecosystem services, and more broadly for urban studies or urban ecological modelling of the São Paulo Metropolitan Region.

Monitoring changes in tree cover for assessment of deforestation is a premise for policies to reduce carbon emission in the tropics. Here, a U-net deep learning model was used to map monthly tropical tree cover in the Brazilian state of Mato Grosso between 2015 and 2021 using 5 m spatial resolution Planet NICFI satellite images. The accuracy of the tree cover model was extremely high, with an F1-score >0.98, further confirmed by an independent LiDAR validation showing that 95% of tree cover pixels had a height >5 m while 98% of non-tree cover pixels had a height <5 m. The biannual map of deforestation was then built from the monthly tree cover map. The deforestation map showed relatively consistent agreement with the official deforestation map from Brazil (67.2%) but deviated significantly from Global Forest Change (GFC)’s year of forest loss, showing that our product is closest to the product made by visual interpretation. Finally, we estimated that 14.8% of Mato Grosso’s total area had undergone clear-cut logging between 2015 and 2021, and that deforestation was increasing, with December 2021, the last date, being the highest. High-resolution imagery from Planet NICFI in conjunction with deep learning techniques can significantly improve the mapping of deforestation extent in tropical regions.

Deep learning self-supervised algorithms that can segment an image in a fixed number of hard clusters such as the k-means algorithm and with an end-to-end deep learning approach are still lacking. Here, we introduce the k-textures algorithm which provides self-supervised segmentation of a 4-band image (RGB-NIR) for a k number of classes. An example of its application on high-resolution Planet satellite imagery is given. Our algorithm shows that discrete search is feasible using convolutional neural networks (CNN) and gradient descent. The model detects k hard clustering classes represented in the model as k discrete binary masks and their associated k independently generated textures, which combined are a simulation of the original image. The similarity loss is the mean squared error between the features of the original and the simulated image, both extracted from the penultimate convolutional block of Keras “imagenet” pre-trained VGG-16 model and a custom feature extractor made with Planet data. The main advances of the k-textures model are: first, the k discrete binary masks are obtained inside the model using gradient descent. The model allows for the generation of discrete binary masks using a novel method using a hard sigmoid activation function. Second, it provides hard clustering classes–each pixel has only one class. Finally, in comparison to k-means, where each pixel is considered independently, here, contextual information is also considered and each class is not associated only with similar values in the color channels but with a texture. Our approach is designed to ease the production of training samples for satellite image segmentation and the k-textures architecture could be adapted to support different numbers of bands and for more complex self-segmentation tasks, such as object self-segmentation. The model codes and weights are available at https://doi.org/10.5281/zenodo.6359859.