Landuse Landcover Change Detection Using Geospatial Semantic Segmentation U-Net-ResNet Architecture

Landuse Landcover (LULC) change detection is essential for environmental applications. In this research, the aim is to develop a Fully Convolutional Network (FCN) framework for multispectral spatial resolution satellite imagery to detect pixel level LULC features and analyze changes in landcover classes of rock, vegetation, water area, built-up, and farm-land with high accuracy. Deep learning image semantic segmentation approach and Landsat-8-OLI multispectral spatial resolution camouflage s atellite image dataset downloaded from https://earthexplorer.usgs.gov of Jaipur city-Rajasthan-India at latitude 26.9124° N, longitude 75.7873° E for five years (2017–2021) is being used to accomplish the work. A fully convolutional encoder-decoder network is extended by U-Net modified architecture with the ResNet34 to detect LULC features. U-Net architecture is modified for semantic segmentation and ResNet34 helps in the encoding and re-sampling itself. The image and the masked image of the image are divided into tiles consisting of size 32 × 32 pixels with 7-color band channels. Multispectral spatial resolution satellite Landsat-8-OLI camouflage i mage o f Jaipur city in blocks of
$\mathbf{h}\times \mathbf{w}\times \mathbf{n}_{channels}$
dimensions is classified at pixel level using U-Net to produce a classified output of h × w ×
$\mathbf{n}_{classes}$
dimensions. The proposed segmentation approach model offers a solution for mapping Landsat-8-OLI satellite imagery LULC features at the pixel level in five classes with F1-score of Rock-0.6234, Vegetation-0.8943, Water Body-0.8219, Built-up-0.6120, and Farmland-0.7120 and mean segmentation score is 0.7327. IOU Score is also achieved with high accuracy for five landcover classes Rock-0.5568, Vegetation-0.8574, Water Body-0.9078, Built-up-0.7234, and Farmland-0.8280, and the mean segmentation score is 0.7749. This process can also be applied to another image dataset to segment the landcover classes at the pixel level.