Hyperspectral Image Based Rainfall with Humidity Prediction by Feature Extraction and Classification Using Deep Learning Techniques

Climate studies require knowledge of the temperature as well as humidity profiles in atmospheric boundary layer. One benefit of ground-based infrared hyperspectral spectrometer is its ability to measure atmospheric radiances with a moderate vertical resolution and a high temporal resolution. This paper makes use of ground-based infrared hyperspectral measurements to retrieve temperature as well as humidity profiles. Existing inversion methods still have issues with high computational complexity or subpar performance, despite some advancements based on statistical learning or physical models. This research proposes a novel technique for hyperspectral image-based rainfall and air humidity prediction utilizing Deep Learning (DL) model. Input is collected as hyperspectral image, which are processed for noise removal as well as normalization. Then, the image features are extracted using Convolutional Gaussian Vector Multilayer Neural Networks (CGVMNN). This extracted feature has been classified using Transfer Graph Temporal Logistic Neural Network (TGTLNN). the classified output shows region with rainfall and air humidity. the experimental analysis has been carried out for various hyperspectral image dataset in terms of accuracy, average precision, recall, F-1 score, ROC. The proposed technique accuracy 96%, Fl Score of 95%, Average Precision of 94%, Recall of 92%, ROC of 93%.

I. Introduction

The status of the atmosphere can be inferred from a variety of observations made both in situ and remotely by weather observation. Surface observations come from in situ sensors that detect temperature, humidity, and pressure directly. On the other hand, a lot of remote sensing data comes from radar and satellites, which measure radiance as well as reflectivity across distance. In the past, weather phenomena from the past or the current atmospheric condition have been examined through observations. Though they have the potential to greatly improve predicting performance, recent developments in DL approaches offer data-driven weather forecasting utilising meteorological observations [1]. Weather forecasting using deep learning algorithms is a major research topic in the computer vision and weather and climate communities. The well-known Conv-LSTM architecture was created to anticipate future precipitation utilizing radar observations in Hong Kong area and is used in a variety of image prediction applications. As a result, there are numerous studies pertaining to weather forecasting utilising DL methodologies. But even with this improvement, quantitative precipitation prediction remains difficult since good precipitation prediction requires taking into account physical processes of clouds as well as precipitation from particle creation at microscale to precipitation method at synoptic scale [2]. The poor representational resolution of model simulations and observations make it challenging to comprehend clouds and precipitation. Based on their physical properties, various cloud and precipitation microphysics parameterization methodologies assess clouds as well as precipitation processes in numerical weather forecasting method [3]. Due to physical process, most numerical models have limitations when it comes to projecting cloud and precipitation in 1–3 hours. For this reason, radar observation is extensively utilized for nowcasting based on extrapolation approach. Nevertheless, the precipitation system's lifetime is not taken into account by extrapolation techniques. Developing a forecast model for accurate rainfall is one of the biggest challenges facing academics in a variety of domains, including functional hydrology, ecological machine learning, climate data mining, and numerical prediction. Climate prediction is unique among all countries in the globe in terms of the benefits and services provided by the meteorological department. Rainfall anticipation is important because heavy and intermittent rains can have a variety of negative effects, such as destroying crops and farms and damaging property. Therefore, having a better prediction method is essential for early warning that can reduce risks to people as well as property as well as also improve farming practices. Predicting when it will rain is a challenging task with very accurate results. Many devices are available that use various aspects of the climate, such as pressure, temperature, and humidity, to predict when it will rain. By using