top of page

AI x Forest Survey: Exploring Machine Learning Applications for UAV Images

Forest, Photo by Geran de Klerk on Unsplash

To make management decisions, forest managers need to have appropriate information about forest resources to serve as the basis for subsequent cultivation and overall management planning. In recent years, drone technology has flourished, and because of its ability to take high-resolution images, it is becoming an important tool for assessing forest conditions and changes.

However, the forest is large and complex, and while drones have solved the problem of data collection, it is still generally a manual process to identify and analyze images, which takes too much time and is too costly.

Fortunately, as the application of machine learning becomes more widespread, it is also beginning to be applied in the forestry industry, as it can classify and analyze images in an automated way with expert-level recognition technology, thus reducing the work time of a large number of investigations.

What is artificial intelligence?

Artificial intelligence is the use of the high-speed computing capabilities of computers to handle tasks that would otherwise be performed by humans; in short, it is the use of computers to perform the process of human intelligence, so that machines, after being programmed, can operate with intelligence similar to that of humans.

What is machine learning?

Machine learning is the process of classifying the collected data through algorithms and training predictive models so that when new data are available, predictions can be made through the trained models.

When forest survey meets UAV

forest survey
Forest survey (Image source:

The most traditional forest survey method is manual field measurement, which requires people to go to the site to observe and collect data with eyes and instruments, which is not only time-consuming and labor-intensive but also can only obtain small area data due to the field of vision.

Therefore, the application of satellite images or the use of a flying vehicle like a helicopter with camera equipment for measurement has been proposed, but it is easily obstructed by clouds due to the flight height.

In the end, the lower cost, lightweight and sensitive drones that can fly at low altitudes under the clouds are favored to make up for the shortage of telemetry images and general aerial photography and can reduce the cost of acquiring three-dimensional data such as point clouds and become mainstream.

UAV x Machine learning

UAVs can be mounted with either a general lens or LiDAR to take pictures. LiDAR is a rapidly developing active remote sensing technology in recent years, which is very suitable for the estimation of forest texture characteristics and canopy structure. However, the cost is too high to be widely used in forestry surveys. Combining drone images with machine learning is the best approach to maximize the benefits.

Machine learning is a proven method for detecting and identifying objects in RGB images, but it has only recently been widely applied to plant detection. Although many individual algorithms such as tree segmentation have been proposed in the past, the development and testing of these algorithms are usually limited to specific sample stations, but few methods can evaluate data from multiple forest types.

We used the DeepForest algorithm to construct a neural network-based model for identifying each tree in the RGB image.

Why DeepForest?

DeepForest uses a deep learning object detection network to predict the bounding box corresponding to each tree in the RGB image. DeepForest is built on top of the RetinaNet model in the Torchvision envelope and aims to simplify the training model for tree detection.

DeepFroest model
DeepFroest model (Image source:

Given the diversity of tree appearance worldwide, defining a single model for tree canopy is challenging. To address this issue, DeepForest provides an explicit training method that allows users to determine what level of accuracy is needed, then annotate the data and retrain the model to have predictions with sufficient accuracy.

Our works

We used python and machine learning to perform tree counting from drone orthophotos. The scope of this application was taken on June 21, 2022, in Caotun Township, Nantou County, Taiwan. The forest phase is a miscellaneous tree forest, including Litchi trees, Dimocarpus longan trees, various fruit trees, etc. The amount of trees identified is presented in a program.

Example of photography in Caotun
Example of photography in Caotun

Tree counting
Counting the tree numbers from DeepForest result

Analyze the value of applications to provide higher-quality services

In summary, we can see that the benefits of UAV image analysis are huge in this wave of artificial intelligence. In addition to forest survey, it can also be applied to a wide range of information, such as the most current carbon rights and carbon sinks, e.g., how many trees are needed to offset the most current carbon rights and emissions?

With the application of Python and machine learning, we can solve the currently intractable problem of counting trees, and thus significantly improve the speed, scale, and cost of biodiversity and forestry surveys.


For image processing and data transformation, DataXquad meets all your needs!

DataXquad is a Pay-Per-Use online image data transformation service platform that can meet all procedures without use costs. We simplify the most complicated part of image data transformation so that more image data can be utilized and more industry chains can derive value from it.


Individual Tree-Crown Detection in RGB Imagery Using Semi-Supervised Deep Learning Neural Networks

Cross-site learning in deep learning RGB tree crown detection

bottom of page