Statistical approach to gap free time series data generation for remote sense images using Google earth engine

Abstract

Remote sensing is gaining popularity in agriculture systems due to the advancements in technology and the precise spatial and temporal data sensed by Earth Observation satellites, aid in land and water monitoring. But to align with Sustainable Development Goal, SDG 2.0 (Zero Hunger), accurate crop yield forecast is necessary to handle the crop shortages and crop surpluses accordingly. As crop yield prediction depends on efficient crop maps, reliable crop maps (of late mostly done by Machine learning/Deep learning models) require continuous time series data of croplands. Atmospheric attenuation has a big role to play in Optical satellites. Snow / cloud cover / aerosols degrade the onboard recorded values considerably. Cloudy pixel values make the analysis less accurate. Since optical satellites fail to deliver gap free time series data, a technique to recover/predict the missing pixel data becomes essential. To handle such pixel values, a technique to reconstruct or predict the missing values is much needed as Machine learning / Deep learning models are efficient when fed with gap free data. Observations from the conducted study by employing statistical methods over cloud computing platform have revealed an acceptable range of Root Mean Square Error and F1 score. In this paper, investigations done to get a best estimate of the missing pixel values, using a statistical approach is discussed. Rolling statistics (Moving Average), Gap filling, SG (Savitzky Golay) filtering, Interpolation are performed with Normalized Difference Vegetation Index (NDVI) values, over a period of four years at a sugarcane farmland region. Using GEE (Google Earth Engine) as a cloud computing service and satellite dataset provider, NDVI values obtained are compared with the actual ground values. SG filtering gave better approximations of missing values compared to other statistical methods. Our study demonstrates an effective technique in generating gap free data that improvises the performance of crop yield prediction models.