Machine Learning Approach for Land Use Classification of Satellite Images

Abstract

Satellite images that include multispectral and color layers may provide newlineinformation in excess on the region of interest (ROI). Because every layer has newlineits own unique geographical characteristics, researchers are in a position to newlinerecognize different locations all around the world. Many machine learning (ML) newlinealgorithms are prone to differentiate between various types of crops, soil types, newlineland cover types, and other such categories. Researchers often use deep learning newlineclassifiers such as Q-learning, RNN (Recurrent Neural Network), FCNN (Fully newlineConnected Neural Network), etc., despite the heterogeneity in their newlineapplicability, complexity, implementation costs, and performance. Such types of newlinenetworks are too much all-purpose to be used for application-specific satellite newlineimagery classification due to their extensive breadth of coverage. This study newlineproposes an improved deep learning model for classifying region-based satellite newlineimages in real-time in an environment optimized for the particular application. newlineELSET can gather large-scale temporal datasets by making use of Google Earth newlineEngine. These datasets are then subjected to application-specific segmentation newlinemodel filters in concern to evade outliers. The process of segmenting image data newlineis carried out with the assistance of enhanced CNN models such as VGGNet 19, newlineResNet V2, and GoogLeNet. newlineMoreover, classification algorithms, such as GoogLeNet for the land and forest newlinecover classification and ResNet V2 for the water bodies and urban areas newlineclassification, are surpassed by VGGNet 19 in terms of the ability to identify the newlinetype of crop grown as well as any crop damage that may have occurred. newlineBecause these techniques are combined, it is now feasible to recognize layered newlineregions in images with a high precision intensity while only requiring moderate newlinemental effort. As a result of these techniques, the recommended model can newlineidentify land types with an accuracy of 95.4%, categorize urban and water cover newlinewith an accuracy of 96.5%, and recognize crop types with an accuracy of newline98.9%. The proposed model was investigated at various times and locations, newlineix newlineand the findings were consistently accurate throughout. In the study, accuracy, newlinerecall, area under the curve (AUC), and latency (delay) were tested against newlinethose of other methods currently considered to be state-of-the-art. newlineCompared to the existing models, the recommended model had an accuracy of newline5% higher, a recall of 8% higher, and an area under the curve that was 5.9% newlinehigher. Because it organizes information more nuancedly than the typical newlinealternatives, the suggested method substantially increases the amount of latency newline(delay) experienced. Because of this study, researchers now have access to newlinemethods for identifying context-specific satellite images, which might shorten newlinethis delay without compromising the levels of model performance