Convolution Neural Network-based Agriculture Drought Prediction using Satellite Images

Abstract

Water scarcity due to limited water resources and substantial increase in water consumption leads to drought situation many parts of world. Agricultural drought affects all walks of life in varyingly severe magnitudes. No solutions exist to completely avoid drought conditions. Analyzing past patterns through drought monitoring for early warning and predicting future drought conditionshelp to mitigate or control causes of drought in many sectors such as water management, agriculture, dam utilization and other various ecosystems. Traditionally, agriculture drought monitoring is typically supported by drought indices such as standardized precipitation index (SPI), Multivariate SPI, Palmer drought severity index, normalized difference vegetation index (NDVI), crop moisture index, Standard Precipitation Evapotranspiration Index, Z-index and crop specific index, and effective drought index etc to evaluate drought effects in agriculture sector. Artificial intelligence (AI) models complex and uncertain situation scholarly. Research on drought prediction uses various AI models based on rainfall/temperature data or satellite data. Convolution neural network (CNN) based drought prediction does not exist. The main objective of this paper is to predict drought or no drought condition using CNN model on the basis of NDVI computed from satellite images for a given location in Kolar district of Karnataka, India. Satellite images are obtained from National Remote Sensing Center (NRSC) and converted to NDVI images. The evaluation of CNN model includes Nash index, correlation coefficient and root mean square error. The improved accuracy is observed up to 96%.