Individual Tree Detection in Large-Scale Urban Environments using High-Resolution Multispectral Imagery

TL;DR

This study addresses scalable, automatic detection of individual urban trees across California using high-resolution NAIP multispectral imagery and a point-annotation–driven learning approach. The authors introduce HR-SFANet, an attention-guided, full-resolution confidence-map regression network that localizes trees via peak finding, trained with a Gaussian-ground-truth confidence map and a composite loss. On a Southern California 2020 test set, their method achieves a precision of 0.736, recall of 0.733, and RMSE of 2.16 m, outperforming baselines PyCrown and DeepForest, with further gains from hyperparameter tuning. They demonstrate wall-to-wall California applicability, generalization to other years and climate zones, and practical utility for urban forestry inventories, while outlining future work to incorporate higher-resolution data and additional tree attributes such as size and species.

Abstract

We introduce a novel deep learning method for detection of individual trees in urban environments using high-resolution multispectral aerial imagery. We use a convolutional neural network to regress a confidence map indicating the locations of individual trees, which are localized using a peak finding algorithm. Our method provides complete spatial coverage by detecting trees in both public and private spaces, and can scale to very large areas. We performed a thorough evaluation of our method, supported by a new dataset of over 1,500 images and almost 100,000 tree annotations, covering eight cities, six climate zones, and three image capture years. We trained our model on data from Southern California, and achieved a precision of 73.6% and recall of 73.3% using test data from this region. We generally observed similar precision and slightly lower recall when extrapolating to other California climate zones and image capture dates. We used our method to produce a map of trees in the entire urban forest of California, and estimated the total number of urban trees in California to be about 43.5 million. Our study indicates the potential for deep learning methods to support future urban forestry studies at unprecedented scales.

Figures (8)

• Figure 1: Locations of cities from which we collected and annotated images to form our dataset. Each climate zone in California is represented by at least one city in the dataset. The climate zones represent the following proportions of the total urban reserve area across the state: Inland Empire (24.03%), Inland Valleys (26.36%). Interior West (2.25%), Northern California Coast (16.46%), Southern California Coast (24.65%), and Southwest Desert (6.25%).
• Figure 2: Example of tree detection using our method on a section of 2020 NAIP imagery in Santa Monica, CA. We process the input raster (a) in a CNN to produce a confidence map (b). We then apply peak finding in the confidence map to produce individual tree detections (c).
• Figure 3: HR-SFANet network architecture. The input image is encoded through the first five blocks of convolutional and pooling layers from the VGG-16 network. Separate attention and confidence heads upsample and aggregate the outputs of the backbone layers to produce an attention map and a confidence map. These are multiplied together to produce the final confidence map.
• Figure 4: Example results from our tree detection method on images from Southern California 2020 test set. The method is able to detect and accurately localize most of the trees in the images. Some areas contain examples of missed detections of trees with small canopies and shrubs being confused with trees.
• Figure 5: Analysis of test set performance with increasing training set size.