Time Series Classification using Improved Deep Learning Approach for Agriculture Field Mapping

Abstract

Agriculture holds an important role in food security management, hence providing the authorities with reliable and updated agriculture field maps from regional to national scale is critical. Unfortunately, conventional digitation on the screen is still dominating the process of mapping production. The recent advancement in remote sensing research has made it possible to optimize the operation of mapping by employing Deep Learning (DL) algorithm to automate the process. This study implemented a novel DL architecture based on multiple blocks of CNN layers which are complemented by a Bi-LSTM and dual FCN layers. Time-series datasets of NDVI were extracted from Landsat 8 OLI (Operational Land Image) ranging from May 2013 to September 2021 as the main features. The validation accuracy score of our DL model during the fitting process was 0.9833. MSAVI replaced NDVI as part of the experiments and our model produced a validation accuracy score of 0.9667. In the latter stage of the experiment, we produced the final comparison using IoU metrics between prediction maps of the agriculture field from our model, ResNet, and ESA WorldCover. Prediction maps from our model topped the chart with highest IoU score amongst others for the NDVI and MSAVI datasets