Artificial Intelligence-Driven Detection of Forest Stands Utilizing Satellite Remote Sensing Data

1A1R RAJA KUMAR, 2HARNALLI ABDUL MOIN, 3MOVVA SWETHA SREE,4DIRASANTHAM SAI PREM,5LANKOTHU JAYADEEP KUMAR

ssistant Professor, Department of Information Technology, SV college of Engineering, Tirupati, India

2B.Tech, Department of Information Technology, SV college of Engineering, Tirupati, India

3B.Tech, Department of Information Technology, SV college of Engineering, Tirupati, India

4B.Tech, Department of Information Technology, SV college of Engineering, Tirupati, India

5B.Tech, Department of Information Technology, SV college of Engineering, Tirupati, India

.Email: rajakumar.r@svce.edu.in, abdulmoin24271@gmail.com, swethamovva35@gmail.com, dirasanthamsaiprem@gmail.com, jayadeep3303@gmail.com

Corresponding Author/Guide: R Raja Kumar, M.tech(Ph.D), Assistant Professor

ABSTRACT:Current Forest stand identification relies heavily on manual field inspections and GIS data entry, which are time-consuming and labor-ntensive. While satellite and aerial imagery have been used to support these efforts, traditional methods—such as bounding box objectdetection—often yield imprecise results and struggle with tasks like tree age estimation. Existing datasets for forest cover detection are typically limited in size, quality, and diversity, especially for buildings and roads, leading to imbalanced training and reduced accuracy. Toaddress these challenges, a new system leverages advanced deep learning techniques, including convolutional neural networks (CNNs)with instance segmentation. The approach uses a custom, high-quality dataset (ForestFullV2) containing over 5,100 labeled images of forests,fields, roads, buildings, and lakes. State-of-the-art models—YOLOv8, YOLOv5, and Mask R-CNN—are evaluated, with the YOLOv8 Small model (batch size 4) achieving the highest accuracy. A novel instance segmentation method enables pixel-level object delineation,significantly improving over traditional bounding box detection. Dataset balancing through augmentation of underrepresented classes(buildings and roads) further enhances model precision. The system automates forest cover monitoring, reduces manual labor and time, and improves the accuracy of forest ecosystem assessments by integrating AI with remote sensing.KEYWORDS: GIS data, object detection, convolutional neural networks, YOLOv8, YOLOv5, and Mask R-CNN, deep learning techniques, augmentation.