Crop Monitoring via Remote Sensing

Abstract

The increasing world population is generating higher food demand that needs effective cultivation methodologies to meet them, especially for the developing countries that have a higher dependence on the agricultural sector. Crop type classification is part of crop monitoring which can help to plan crops effectively and meet the demand supply chain. Our study had mainly two objectives, acquisition of a dataset for crop type classification and building effective models for Pakistan-specific regions that can have comparatively better outcomes for the region. The dataset was acquired from different regions of Pakistan via local surveys and later on perform post-cleaning to get an optimized model especially for LSTM where data was converted into a timeseries dataset which provided us comparatively more accurate results. The dataset had sentinel-2 images ranging from 2016 to 2021 for mainly 5 crops and a no-data class capturing both Kharif and Rabi seasons of the area. We used high temporal and spatial resolution images to train Temp CNN, Light GBM, and LSTM where we achieve a model having an accuracy of 94%. The LSTM model on time-series data outperformed where the spatial and temporal pixel of each location was converted to a time dimension. The developed methodology can be used to forecast the supply of different crops as well as the models can be trained on more crop types. The acquired dataset can be used to try different methodologies for developing optimized models.