39Hydro Utilizing Satellite Imagery and Machine Learning to Quantify Drought Risk Index in Thailand’s Agricultural Regions

Drought is one of the major natural disaster that continuously impacts Thailand's agricultural sector, affecting it during regular dry seasons as well as in the face of unpredictable rainfall patterns. Thailand's agriculture primarily focuses on crops like rice, sugarcane, and cassava. However, these activities have suffered extensive damage and economic losses due to drought conditions caused by climate change. Thirty percent of Thailand's population work in agriculture, and drought threatens their income, livelihood, quality of life and poses food security issues. To this end, a more comprehensive understanding of this phenomenon is a prerequisite for developing drought monitoring and tools from satellite technology and machine learning approach.

This also aligns with the Sustainable Development Goals (SDGs), particularly SDG 2 (Zero Hunger) and SDG 13 (Climate Action), to create a sustainable agricultural system that is resilient to climate change.

However, aiding affected farmers requires substantial resources, including budget, labor, and time, which can hinder timely recovery. Geoinformation technology provides factual and verifiable information for more effective interventions. Droughts' increasing frequency and severity pose significant challenges to agricultural sustainability and food security. The National Research Council of Thailand (NRCT) funded this research project during the fiscal year 2023. The research team collaborated with various government departments, including the Department of Agricultural Extension, the Royal Irrigation Department, the Meteorological Department, the Office of National Water Resources, the Department of Disaster Prevention and Mitigation, and the Office of the Permanent Secretary for Agriculture and Cooperatives. They utilized data from these departments to conduct a comprehensive study, analysis, and assessment of areas at risk of drought. At present, there is lack efficient system in the region to analyze and deliver drought-related information to the stakeholders on the ground. The objectives of this study are to quantify Drought Risk Index using satellite imagery and machine learning.

This study is to create a Drought Risk Index (DRI) assessment model by incorporating various factors including the Standardized Precipitation Index (SPI), Soil Moisture (SSM), Vegetation Health Index (VHI), Land Use Data (Level_03), Soil Type Data, Groundwater Storage (GWS), and Evapotranspiration (ET). This involves integrating data from drought-affected and non-drought-affected plots between 2015 and 2022. Comprehensive data preparation was conducted, including cleaning, standardization, and parameter setting on a dataset comprising 96,687 drought-damaged plots and 5,615,607 non-damaged plots. The machine learning algorithm underwent training to comprehend the attributes of indices associated with drought, the typical growth of plants, and previous instances of damage. This training was conducted utilizing data from plots that were either unaffected or harmed, encompassing four crop varieties: rice, maize, cassava, and sugarcane. A predictive model for drought-risk areas was developed using the Random Forest classification approach. The model focused on nine provinces: Kamphaeng Phet, Uthai Thani, Nakhon Ratchasima, Surin, Roi Et, Sakon Nakhon, Nakhon Sawan, Phichit, and Suphan Buri.

The model was created with an automated drought area assessment technique. The model was trained using a Random Forest classification, achieving an accuracy score of 92%, a precision of 92.5%, a recall of 92%, and an F1 score of 91.9%. To validate the model, The model's predictions were validated against field-surveyed soil moisture data from 9 provinces and satellite data from Soil Moisture Active Passive (SMAP) and Moderate Resolution Imaging Spectroradiometer (MODIS). The study's results reveal a strong relationship (R² = 0.75) between field-surveyed soil moisture and the Drought Risk Index, indicating the model's robustness in practical applications. As results, we develop mobile application that provide real-time alerts such as Drought Risk Index (DRI), Soil moisture etc. With the app, users can receive alerts on drought risks, weather forecasting, and Drought Risk forecasting. A free version of the mobile app, ' Check Drought ', is available for users on both iOS and Android platforms and the website http://cropsdrought.gistda.or.th. The overall aim of this study was provides valuable tools for policymakers and farmers to predict and manage drought risks, enhancing Thailand's agricultural resilience and food security.

Keywords: random forest, machine learning, remote sensing, vegetation indices, drought risk index