In the realm of tree species detection and classification from remote sensing data, the authors of the Tree Species Detection dataset highlight the domination of multispectral and hyperspectral images, along with Light Detection And Ranging (LiDAR) data. Notably, despite the availability of Red-Green-Blue (RGB) sensors, only a limited number of studies have delved into the utilization of RGB sensor-captured images. Furthermore, the emergence of deep learning techniques for object detection in recent years has introduced significant advancements. In light of these developments, the authors proposed and assessed the application of Convolutional Neural Network (CNN)-based methods, in conjunction with high spatial resolution RGB imagery obtained from Unmanned Aerial Vehicles (UAVs), for the purpose of detecting tree species under legal protection.

The preservation of sensitive tree species necessitates timely and precise information regarding their distribution in endangered areas. Remote sensing techniques have gained traction as alternatives to labor-intensive and costly field surveys for assessing forest resources. These techniques encompass the use of satellites, aerial surveys, and, more recently, Unmanned Aerial Vehicles (UAVs) as the primary platforms for data collection.

To the best of the authors’ knowledge, no prior research has centered on the application of state-of-the-art deep learning-based methodologies for tree detection using images acquired from RGB cameras mounted on UAVs. The authors aim to bridge this gap and offer an evaluation of deep learning-based approaches for the individual detection of trees in high-resolution UAV RGB imagery. Their study specifically concentrates on the tree species Dipteryx alata Vogel (Fabaceae), commonly referred to as “baru” or cumbaru. This species faces a significant risk of extinction, as indicated by the IUCN (2019) (The International Union for Conservation of Nature’s Red List of Threatened Species). Cumbaru trees hold substantial economic potential and find diverse applications, primarily in the realm of non-timber forest products. They exhibit a wide distribution, predominantly within the Brazilian Savanna, although occurrences have been noted in the wetlands of Midwest Brazil.

The experimental process unfolded in four primary phases. Firstly, RGB images were acquired at various points throughout the year using a UAV-based camera. A specialist was tasked with annotating these images by delineating cumbaru trees using bounding boxes. Subsequently, the networks associated with each method were trained to detect instances of cumbaru trees within a designated set of training images. In the third step, the trained networks were applied to images not included in the training dataset, with the objective of predicting the presence of cumbaru trees and producing corresponding bounding boxes. The final phase involved the computation of accuracy metrics for each method based on the predicted outcomes.

A total of 392 UAV images (Please note, that DatasetNinja holds an updated version of the dataset with 712 images) were procured over a seven-month period, spanning from August 2018 to February 2019, across six missions. The advanced Phantom 4 UAV, equipped with a 20-megapixel camera, captured these images at flight altitudes ranging from 20 to 40 meters (refer to Table 1 for comprehensive details). These images exhibited a mean Ground Sample Distance (GSD) of 0.82 cm (centimeters) and were acquired within three distinct study regions, encompassing a total area of approximately 150,000.00 square meters, situated in the urban domain of Campo Grande municipality, Mato Grosso do Sul, Brazil. These missions collectively encompassed around 110 trees, each captured with variations in appearance, scale, and illumination:

Examples from the dataset: images captured on: (a) 26 August 2018; (b) 21 September 2018; ( c) 22 September 2018; and (d) 20 February 2019

Each image within the dataset underwent meticulous annotation by a specialist employing LabelMe software. During this annotation process, a bounding box delineated by the top-left and bottom-right corners was assigned to each cumbaru tree specimen present in the image.