Aerial Imagery Segmentation of Landcover by Using Deep Learning Algorithm

An essential task in geospatial analysis is accurately segmenting land cover from high-resolution aerial images, which supports applications including the management of resources, urban development, and ecological monitoring. This research introduces a deep learning-based framework utilizing the DeepLabV3+ architecture with a ResNet50 backbone to achieve precise pixel-wise semantic segmentation of land cover types. The landcover.ai dataset, which offers high- resolution aerial images and correlating segmentation masks that represent different land cover classes, is used in the research. To standardize input dimensions, huge photos are split into <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$256 \times 256$</tex> patches as part of the methodology's thorough data preparation. To improve model generalization, data pretreatment procedures include MinMax normalization as well as one-hot encoding for multi-class segmentation masks, and reliable data augmentation methods including randomized flips, and rotations The DeepLabV3+ low-level maps of features are used in the model to provide precise boundary segmentation and uses Atrous Spatial Pyramid Pooling (ASPP) to collect multi- scale context-relevant data. The training procedure uses a loss function hybrid that mixes Focal Loss and Jaccard Loss to handle the class imbalance. The Adam optimizer is used for improving the model, and ReduceLROnPlateau is used to dynamically alter the learning rate. Model performance is evaluated using a testing dataset using evaluation measures like Mean Intersection over Union and Pixel Accuracy and then post segmentation process with smooth blending strategy The segmentation results are computed by taking the class-wise maximum probabilities from the softmax output using argmax. The processed predictions and original mask are saved for further analysis. The ability of sophisticated neural network designs, such as DeepLabV3+, to analyze highresolution aerial imagery is demonstrated by this research. The suggested method provides a scalable and the effective way to map land cover which helps with the ecologically friendly growth, urban resource development, and better ecological monitoring.