Feature selection-based hybrid DL model for detecting the unwanted land using satellite data

Almost every modern application (including farming, logging, and mining) that requires observed data from above or sideways perspectives makes use of UAV sensors and platforms. Despite their best efforts to become a universal remote sensing (RS) instrument, most of the techniques for processing and analysing RS data are still case-specific. For the purpose of weed and crop classification using UAV photos, this research investigates the feasibility of using hybrid deep learning algorithms. Images, feature extraction, and labelling of photos have been used to train deep learning systems for the difficult job of weed detection in crops. A deep learning approach using correlation filtering is offered as a solution to the issues surrounding the reliable identification of visual objects. This research suggests using (RPReLU) as the activation function in conventional (CNN) to address the issues of mean deviation and inadequate discriminate capacity. However, the typical loafer process’s zero-setting operation of weights is random, so it can’t tell the difference between features with different weights and hence has a poor learning productivity. As a result, the authors of this paper present a dropout strategy based on a (SVM) that allows for a more selective dropout rate, thereby enhancing the learning efficiency and manual direction of the dropout process. To add insult to injury, conventional CNN only use the output of the last layer, which is great for capturing semantic information but not spatial ones. We recommend using the many convolution layers in CaffeNet as the target representation because to their rich features and ability to overcome this issue. In addition, an IHGS algorithm is created to choose the characteristics of UAV photos best by boosting the hunger games search’s intensification and diversification capabilities (HGS). Performance was compared using accuracy, precision, recall, and the F-measure.