Mapping the vegetative stage of rice cultivation using deep learning

Monitoring rice growth stages is important for maintaining food security, especially from flood and drought, which is more frequent due to climate change. The Sentine1-2 Rice Growth Stages (S-2 RGS) monitoring project from the Ministry of Agriculture is one of the ongoing monitoring systems (http://katam.litbang.pertanian.go.id/sc). However, S-2 RGS has difficulty providing a cloud-free project, especially in the wet season. Sentinel-1 (S-1) has recently been used to identify mapping rice area due to better spatial resolution than the MODIS dataset. This study focuses on mapping the vegetative stage using a multi-look of S-1 dataset with a deep learning method. Deep learning has been proven to be one of the classification methods on non-parametric approaches that give high accuracy, high computation speed, and less memory usage. The method of this study is as follows: First, the S-1 images downloaded from Google Earth Engine, which has been preprocessed accordingly. Secondly, the rice growth stages condition of Indramayu Regency was extracted from S-2 RGS using random sampling on 16 May 2020. Next step, both multi-look S-1 (Ascending, Descending). Third, the values of RGS was reclassified from four classes (Bare land, Vegetative, Reproductive, and Ripening) into two class (vegetative and Non-vegetative). Finally, the vegetative stage was modeled from four predictors (DESC-VH, DESC-VV, ASC-VH, ASC-VV) using deep learning using two layers with a training dataset (n=9,319). The accuracy of the model is 73.6% (n=3,994). This result shows that deep learning can predict a vegetative stage map with high accuracy where it can fill a cloudy area in the RGS-2 system to provide a cloud-free rice growth stages map in near real-time.