Design of a High-Efficiency Temporal Engine for Real-Time Spatial Satellite Image Classification Using Augmented Incremental Transfer Learning for Crop Analysis

Satellite image processing is a multidomain task which involves design of image capturing, denoising, segmentation, feature extraction, feature reduction, classification, and post-processing tasks. A wide variety of satellite image processing models are proposed by researchers, and each of them has different data and process requirements. For instance, the image capturing module might obtain images in layered form, while feature extraction module might require data in 2D or 3D forms. Moreover, performance of these models also varies due to changes in internal process parameters and dataset parameters, which limits their accuracy and scalability when applied to real-time scenarios. To reduce the probability of these limitations, a novel high-efficiency temporal engine for real-time satellite image classification using augmented incremental transfer learning is proposed and discussed in this text. The model initially captures real-time satellite data using Google’s Earth Engine and processes it using a transfer learning-based convolutional neural network (CNN) via backscatter coefficient analysis. These coefficients indicate average intensity value of Precision Image (PRI) when evaluated over a distributed target. Due to extraction of backscattering coefficients, the model is capable of representing crop images in VV (vertical transmit, vertical receive), and HV (horizontal transmit vertical receive) modes. Thereby assisting the CNN model to extract a wide variety of features from input satellite image, which classifies these datasets (original, VV, and VH) into different crop categories. The classified images are further processed via an incremental learning layer, which assists in visual identification of affected regions. Due to use of incremental learning and CNN for classification, the proposed TRSAITL model is capable of achieving an average accuracy of 97.8% for crop type and severity of damage detection, with an average PSNR (Peak Signal-to-Noise Ratio) of 29.6 dB for different image types. The model was tested on different regions around our local geographical area, and consistent performance was observed. This performance was also compared with various state-of-the-art approaches. It was observed that the proposed TRSATL model has 5% better accuracy, 4.6% better precision, and 7.9% better recall when compared to them, which makes it highly useful for real-time satellite-based crop classification applications.