Semantic Segmentation of Satellite Images for Accurate Forest and Crop Delineation Using Deep Learning Technique

Accurately segmenting forests and crops is essential for efficient land management and resource allocation. Traditional methods rely on manual processes, which are labor-intensive and time-consuming. Semantic segmentation offers a solution by categorizing pixels in satellite images, facilitating precise delineation of land cover types. Utilizing a dataset of 800 satellite images from Google Earth, with an 80:20 training and testing split, ground truth images were generated through manual annotation using APEER software. Deep learning methods employing the U-Net architecture with ResNet34 and VGG-19 backbones were employed for efficient segmentation. Evaluation metrics included accuracy, precision, and recall measures. The U-Net model achieved an accuracy of 95.15%, while incorporating ResNet34 and VGG-19 backbones yielded accuracies of 98.36% and 71.33%, respectively. Crop and forest area estimation was conducted based on pixel counts within regions of interest. Comparative analysis of segmentation metrics demonstrated the model’s performance against original images. The study underscores the effectiveness of deep learning methods, particularly U-Net with ResNet34 backbone, in accurately segmenting forest and crop areas in satellite imagery.