Prediction of Rainfall Categories in Kerala, India, Using Satellite Imagery and a YOLOv8 Classifier

India is a geographically diverse country with a variety of climatic regions that have an important effect on its agricultural methods and natural landscapes. The nation can be largely classified into five main climate zones: tropical, arid, semi-arid, temperate, and alpine. Kerala is located at the southwestern corner of India, with the Arabian Sea on its western side. It shares borders with two other Indian states: Karnataka to the north and northeast, and Tamil Nadu to the east [1]. Kerala state of India lies between the northern latitude of 8°.17'.30" N and 12°.47'.40" N and east longitudes 74°.27'.47" E and 77°.37'.12" E [2]. According to the Köppen climate classification, Kerala state falls under the tropical monsoon climate type [3]. Kerala receives an average rainfall of more than 3000mm, and is mostly obtained from the seasonal monsoon. Unlike in the past years, Kerala has been witnessing severe floods during recent years. The flood of 2018 has been named as the “flood of the century”. It has claimed over 400 lives, affecting the lives of millions of people. The Indian Meteorological Department (IMD) has classified the rain received during those days as “large excess”. Subsequently, in 2019 and 2020 also, the state has witnessed floods and landslides. Kerala is experiencing significant fluctuations and unpredictability in its weather conditions, particularly in the timing and volume of rainfall. In this context, accurate and timely prediction of meteorological parameters and daily rainfall has got broad implications from enhancing crop yields to minimizing casualties and property destruction.

The majority of early techniques for forecasting daily rainfall relied on statistical models to find trends and connections in historical meteorological data. Time Series Analysis can accurately simulate the temporal dependencies found in rainfall data, methods such as Autoregressive Integrated Moving Average (ARIMA) models is widely used. These techniques were developed by Box and Jenkins [4] and have been extensively utilized in meteorology. Regression models, both linear and nonlinear, are frequently used to forecast rainfall based on wind speed, temperature, and humidity. For example, multiple regression models were used by Kidson and Thompson [5] to enhance short-term weather forecasts. Complex patterns in rainfall data can be captured with tremendous potential by neural networks, particularly architectures like Long Short-Term Memory (LSTM) networks. For instance, Shi et al. [6] used convolutional LSTM networks to show how well deep learning works for precipitation prediction. Prediction accuracy can be increased by combining machine learning techniques with conventional statistical methods. The combination of Singular Spectrum Analysis (SSA), Auto Regressive Integrated Moving Average (ARIMA), and Artificial Neural Networks (ANN) into a hybrid model (SSA-ARIMA-ANN) for generating reliable daily rainfall forecasts in a river catchment [7]. Ridwan et.al. has forecasted rainfall using two approaches: (1) Autocorrelation Function (ACF) based on past rainfall data, and (2) Projected Error based on both past and projected rainfall data. Various algorithms, including Bayesian Linear Regression, Boosted Decision Tree Regression, Decision Forest Regression, and Neural Network Regression, were employed to identify the optimal predictions for rainfall across different time horizons. The results showed improved accuracy when cross-validation was applied, particularly with Boosted Decision Tree Regression, after tuning its parameters [8].

Numerous methods using satellite imagery were investigated and tested to reliably predict the daily rainfall. Due to the highly unpredictable nature of rainfall, sufficient prediction accuracy could not be achieved. Boonyuen et al. [9] used the satellite images of Thailand and categorized images into “Rain” and “No Rain” categories. They used the Convolutional Neural Network method to predict rainfall category and obtained prediction accuracies of about 77.42% for current day rainfall and one day ahead rainfall, 67.47% for two days ahead and 58.06% for three days ahead prediction. Novitasari et al. [10] have used Himawari-8 IR satellite imagery and meteorological data. Adaptive average brightness thresholding method was employed for segmentation and classification was done using back propagation. Simanjuntak et al. [11] used high spatiotemporal Himawari 8 satellite imagery and employed the multivariate LSTM forecasting method for rainfall prediction. Ionescu et al. [12] proposed DeePsat, a CNN model that predicted future satellite images from the past satellite imagery. Liyew and Melese [13] used meteorological data of Ethiopia to analyze the rainfall data. The Machine learning algorithms used in this work are: Multivariate linear regression, Random Forest and XGBoost. Research work was published for forecasting rainfall on yearly, monthly, daily and even on an hourly basis [14, 15].

From a layman perspective, a linguistic description of rainfall gives more impact rather than the exact quantity of rainfall. This paper aims to predict the category of daily rainfall. In this work the daily rainfall is categorized as: No Rain (NR), Very Light Rain to Light Rain (VLR-LR), and Medium Rain to Heavy Rain (MR-HR). The methodology adapted is the YOLOv8 classifier. YOLOv8 is the recent version of real-time object detection and image segmentation model [16]. YOLO's popularity stems from its ability to achieve a high-level of accuracy while maintaining a compact model size. YOLOv8 leverages state-of-the-art innovations in deep learning and computer vision technologies, delivering exceptional results with regard to both processing speed and detection precision. The work is focused on the YOLOv8 model for rainfall category prediction. For a comprehensive evaluation of its effectiveness YOLOv8 models are compared with ResNet-50, an established deep learning architecture for image classification. This comparison will establish the effectiveness of YOLOv8 models in predicting the rainfall category.

The remaining part of the paper is organized as follows. Section 2 investigates the previous works related to daily rainfall prediction and YOLO. Section 3 elaborates on the rainfall pattern in Kerala, India. Section 4 includes related work analysis. Section 5 details on YOLOv8 classifier structure, and finally, Section 6 includes results and discussions of the YOLOv8 models and the comparison with the ResNet-50 model.

In this work, YOLOv8 is used for predicting the category of rainfall. For that, the YOLOv8 is trained using the satellite images of the previous day. Meteosat IODC Airmass images of Southern Asia, obtained from EUMETSAT are used for training, validation and testing of the model NVIDIA GeForce RTX 3050 8GB GPU in the Linux environment used for developing the prediction model.

The following are the main features of the proposed model:

(1) The proposed model utilizes daily rainfall data sourced from the Indian Meteorological Department (IMD) website for classifying days into NR, VLR-LR, and MR-HR categories.

(2) Satellite images sourced from EUMETSAT are employed in the proposed model. Airmass images from the preceding day are utilized for training, enabling the model to predict the next day's rainfall category when tested with the satellite image from the previous day.

(3) The proposed system uses the YOLOv8 classifier to achieve high prediction accuracy. There is no existing literature that reports using YOLOv8 for the prediction of daily rainfall category.

YOLOv8 is the latest and cutting edge YOLO approach which can be used for object detection, image classification as well as segmentation [18]. It was released by Ultralytics in the year 2023. The architecture of YOLOv8 builds upon previous versions of YOLO algorithms. This is the first study that has employed the YOLOv8 algorithm for rainfall category prediction using satellite images. The proposed approach provides improved accuracy, versatility and reduced false classification. YOLO has gained popularity due to its ability to achieve a good balance between accuracy and model size [18]. YOLO is designed to detect and classify objects in images and video frames. YOLO uses a grid based approach, where the image is divided into a grid and each grid cell detect objects contained within itself.

The architecture of YOLOv8 consists of two primary components: a backbone network that extracts features and a head network that processes these features for detection tasks [19]. YOLOv8 is an improvement over YOLOv5 that combines the Cross-Stage partial idea [19], feature fusion method [19] and PPF module. The backbone of YOLOv8 is a modified version of CSPDarknet53 (Cross-Stage Partial) architecture. Architecture has 53 convolutional layers and uses cross-stage partial connections [18]. The head of YOLOv8 consists of multiple convolutional layers followed by a series of fully connected layers. An essential enhancement in the YOLOv8 model involves integrating a self-attention mechanism within the network's head [20]. This allows the model to focus on various parts of the image. The Feature pyramid network consisting of multiple layers detects objects at different scales [21]. The different model sizes of YOLOv8 are: YOLOv8-n (nano), YOLOv8-s (small), YOLOv8-m (medium), YOLOv8-l (large) and YOLOv8-xl (extra-large) [20]. The size of the model is proportional to mean average Precision (mAP) and is inversely related to inference time. YOLOv8 classifier identifies and outputs a single class based on the probability values. The result of the classification task includes both a class index and a confidence score.

The proposed model predicts the next day’s rainfall category using the YOLOv8 model. For this, the satellite images are categorized into NR, VLR-LR and MR-HR, depending on the amount of rainfall received on the subsequent day. The YOLOv8 model is trained using these categorized satellite images. YOLOv8 with its deep neural network architecture, performs well in classification tasks. The dataset for the model consists of satellite images obtained from EUMETSAT. Figure 1 shows a sample satellite image. The satellite image has been trimmed to match the dimensions of the area of interest, such that it includes the total of Kerala state and the nearby Indian Ocean region. The 41.5^o Indian Ocean Data coverage (IODC) Southern Asia’s Airmass images are obtained from Meteosat-9 satellite.

1.png

a) Airmass image of Southern Asia

1b.png

b) Cropped image

Figure 1. Airmass and cropped views of Southern Asia

A total of 1900 images were used for the prediction model. A dataset of Airmass images for South Asia has been meticulously compiled from the EUMETSAT website, with daily downloads initiated in 2019 and continuing through 2023. This dataset serves as the resource for the current research work, consisting of a total of 1,900 satellite images. The images are categorized primarily into three groups: the majority fall under the NR category, and the lowest number of images are MR-HR.

The limited number of images in the MR_HR category reflects the variability in precipitation patterns across the region, which can impact data availability. However, the dataset's current size of 1,900 images provided a substantial foundation for the research work. As data collection continues over subsequent years, there is significant potential to expand this dataset further. Increasing the number of satellite images will improve the predictive accuracy of the proposed model. The number of images used for training, validation and testing for each class is depicted in Table 2.

Table 2. Dataset description

The training process for the rainfall prediction model involves utilizing satellite imagery captured on the preceding day as input data. Subsequently, during the testing phase, the model is provided with the satellite image from the prior day as input, and it performs classification to predict the category of rainfall for the current day.

Within the YOLOv8 framework, various model sizes are available, denoted as YOLOv8-n-nano, YOLOv8-s-small, YOLOv8-m-medium, YOLOv8-l-large, and YOLOv8-x-extra-large. The relationship between model size, mean average precision (mAP), and inference time in YOLOv8 is characterized by linear proportionality to mAP and inverse proportionality to inference time [22]. Larger models exhibit higher mAP but demand more inference time for accurate object detection. Conversely, smaller models offer faster inference times but tend to have lower mAP. Opting for bigger models is advantageous when dealing with limited data.

This research trains multiple YOLOv8 variants along with a ResNet-50 model on the customised dataset, followed by comprehensive performance assessment and comparison to determine the optimal model for daily rainfall category prediction. The latest version of YOLOv8 library, YOLOv8.0.232, is used to develop the proposed prediction model. To construct the prediction model, a custom dataset is created, comprising satellite images representing distinct rainfall categories. The dataset creation process involves establishing a "data" folder, within which three subfolders named "train," "test," and "val" are created. Each of these subfolders further encompasses three folders, each designated with the name of a specific rainfall category, containing the respective satellite images. The folder arrangement of the dataset is shown in Figure 2.

Figure 2. Data folder arrangement

The results clearly indicate that all the YOLOv8 models’ average accuracy values are above 90%, with the highest accuracy value of 95% for YOLOv8large. The inferences drawn from the performance comparison chart for each rainfall category are:

NR Category

• YOLOv8 Medium achieved the highest Precision (95%), indicating it has the best ability to avoid false positives in predicting "No Rain."
• YOLOv8 Nano has the highest Recall (96%), implying that it has captured almost all actual "No Rain" instances.
• The F1 Score, which balances Precision and Recall, is highest for YOLOv8 Nano (95%). This makes YOLOv8n the best model for predicting "No Rain".
• ResNet-50 has an F1 Score of only 77%, indicating that its ability to accurately predict No Rain is poor compared to the YOLOv8 variants.

VLR-LR Category

• YOLOv8 Nano shows maximum Precision (89%) and Recall (95%) values, resulting in the highest F1 Score (92%). This means YOLOv8 Nano is the most effective in identifying light rainfall events and minimizes false positives.
• YOLOv8 Large also exhibits better Recall (93%) value, making it capable to detect most light rain cases, but it has a slightly lower F1 Score (88%) due to lower Precision value.
• ResNet-50 has an F1 Score of 76%, indicating a higher rate of misclassification compared to YOLOv8 models

Moderate Rain to Heavy Rain (MR_HR) Category

• YOLOv8 Large achieves the ideal Precision (100%) value for this category, implying that it made no false positive predictions for "Moderate to Heavy Rain."
• Recall is slightly lower (91%), indicating that it has misclassified some true "Moderate to Heavy Rain" instances.
• YOLOv8 Large and Nano have identical F1 Scores (95%), indicating that they excel in detecting and classifying heavy rainfall events.
• ResNet-50 has an F1 Score of only 67%, indicating it frequently misclassifies heavy rainfall cases, likely leading to missed alerts for such important weather events.

Based on the above observations, the following conclusions can be drawn:

1). YOLOv8 models significantly outperform ResNet-50 in predicting rainfall categories across the board. This demonstrates the effectiveness of the YOLOv8 architecture in extracting relevant features from satellite images for weather classification tasks.

2). YOLOv8 Nano emerges as the best overall with the best F1 Score, making it the most reliable model for accurate predictions of rain intensity.

3). YOLOv8 Large excels in detecting Moderate to Heavy Rain with perfect Precision, meaning it is the best choice in avoiding false alarms for heavy rains.

4). ResNet-50 shows weaker performance in all categories, with its highest F1 Score being only 77% for the "No Rain" category.

There is a probability of misclassifying the MR-HR categories as VLR-LR and this poses significant challenges in effectively preparing the public for potential weather-related calamities. Such misclassifications can hinder timely responses and appropriate resource allocation during adverse weather events.

To enhance the model's accuracy, one effective approach would be to increase the volume of satellite imagery included in the dataset. Additionally, integrating more climatic parameters-such as humidity, temperature, and atmospheric pressure-could provide the model with valuable contextual information that enhances its predictive capabilities.

The authors' future research will focus on improving the model's accuracy through these strategies. By expanding the dataset and refining the input features, we aim to reduce misclassification rates and ultimately improve the reliability of rainfall predictions. This will not only enhance the model's performance but also ensure that communities are better prepared to respond to weather-related challenges.

To deploy this model, a Telegram-based application can be developed utilizing the proposed YOLOv8 model for rainfall prediction, allowing the public to access this bot at any time to view daily rainfall category forecasts.