Diogo Nunes Gonçalves's research while affiliated with Federal University of Rio de Janeiro and other places

... This is where unsupervised learning becomes valuable as it can handle such unlabeled data more effectively compared to supervised learning. Semi-supervised learning addresses the issue of inadequate labeled samples by incorporating numerous unlabeled samples along with a small set of labeled ones to train the classifier [15][16][17][18]. Consequently, these algorithms draw knowledge from both labeled M. Mittal et al. and unlabeled data, which adds complexity to the learning process. ...

... While lacking a reference dataset for precise accuracy assessment, these results are in close agreement with the estimation provided by the Forest Protection Department of Lam Dong province, which reported an area of about 13 ha [20], thus suggesting that the methodology used in this study is reliable and accurate, and can be useful to map burned area in near real-time. In comparison with previous work focusing on burned area mapping using deep learning techniques trained on PlanetScope observations [25,26], or the fusion of PlanetScope with other optical satellite observations (i.e., Landsat-8 and Sentinel-2) [2,27], the proposed method is faster and less complicated. This method is suitable for local managers to rapidly generate burned area maps; therefore, it is very useful for emergency response of forest fires, particularly in rural areas. ...

... The traditional method of tree species investigation mainly relies on field surveys, which has the advantages of reliability and accuracy and could provide sample data for remote sensing model training and validation [5]. However, it is time consuming, labor intensive, information limited [6], ...

... Fry counting refers to the counting of the number of targets in a specific area to aid in production decisions [1], [2], [3]. Its accuracy is very important for scientific decisionmaking, including the scientific feeding, behavioral analysis, transportation, and marketing of fish, along with an assessment of fry survival and culture density control [4], [5], [6], [7]. ...

... In addition, deep learning-based approaches, single and two-stage detection techniques are used for object detection. Especially in fruit detection, there are studies in which YOLO [13][14][15] and CNN [16] architectures are frequently used. A detailed comparison of these studies is given in Table 1. ...

... Normally, images input into the Mask R-CNN has limited sizes (e.g., smaller than 1,200 × 1,200 pixel in this study), since too large image size would exceed the memory capacity of graphics cards in the model running. A standard solution to this problem is to split the full UAV images (5,472 × 3,648 pixel) into smaller patches (Osco et al., 2021). Thus, full images in the training dataset were split into image patches (i.e., 1,200 × 1,200 pixel's patches or smaller patches at the image edge). ...

... The first CNNs used for fruit detection were two-stage networks type, with a structure based on two main modules: (1) a region proposal module used to propose ROIs likely to contain a fruit; (2) a classification branch used to classify the proposed regions into fruit or background and refine the detection bounding box. The most commonly used twostage CNN for fruit detection is the Faster-RCNN (Ren et al., 2017), which has been used to detect apples (Apolo-Apolo et al., 2020b;Kang and Chen, 2020;Tian et al., 2019), oranges ( Apolo-Apolo et al., 2020a, Biffi et al., 2021, mangos (Bargoti and Underwood, 2017a;Koirala et al., 2019b), kiwis (Gan et al., 2018), and strawberries , among others. ...

... Our proposed work utilizes the SegNet-based CNN technique [25]. SegNet has two parts: (i) encoder and (ii) decoder. ...

... Schmidt et al. 18 developed a machine-learning model to extract voltage rating information of transmission lines from aerial images. Gomes et al. 19 and Huang et al. 20 used deep learning to localize utility poles and predict line connections from remote sensing images, but the applicability of such methods on distribution grids is subject to varying image resolutions and qualities of remote sensing images across different places, as utility poles and distribution lines, compared with transmission infrastructure, are barely visible from lowresolution or noisy remote sensing images. Deep learning has also been used to either localize or analyze utility poles in street view images [21][22][23][24] , but not for identifying power line connections to construct a full distribution grid map. ...

... Current literature suggests that it is possible with the use of deep learning and tree species classification systems and optical remote sensing techniques, and there is evidence that channels in these algorithms can be substituted with LiDAR metrics [101]. These methods could be applied to CCF stands for stock mapping, mapping of inventory with species distributions and abundance [57,86,87,89,94]; however, for aerial data, occlusion below dense canopy would limit reliability and for terrestrial data, the extent would be limited. Modern ALS methods with laser scanning at angles close to nadir can improve canopy penetration though dense canopy continues to obscure the understory and the close to nadir angled pulses are less likely to reflect off the vertical stem surfaces. ...