3D-CTM: Unsupervised Crop Type Mapping Based on 3-D Convolutional Autoencoder and Satellite Image Time Series

Accurate and timely mapping of different crop types holds significant importance for food security and crop management at regional and global levels. This letter proposes a novel 3-D-crop type mapping (CTM) model based on an unsupervised crop-type mapping technique using satellite image time series (SITS) data and a 3-D convolutional autoencoder (CAE). The study uses a combination of five different vegetation indices: Normalized Difference Vegetation Index (NDVI), Red-Edge Chlorophyll Vegetation Index (RECl), Normalized Difference Red Edge Vegetation Index (NDRE), Green Normalized Difference Vegetation Index (GNDVI), and Visible Atmospherically Resistant Index (VARI) of time series data, along with a 3D-CAE for feature extraction. The model underwent evaluation employing one publicly accessible dataset and one in-house dataset sourced from the PlanetScope instruments. In the case of the in-house dataset, the model has achieved an accuracy of 88.75%, 73.88%, and 93.97% for maize, paddy, and sugarcane, respectively. In the case of the publicly available dataset, the model has achieved an accuracy of 83.71%, 72.81%, and 75.37% for paddy, triticale, and barley, respectively. Our 3D-CTM model demonstrates higher accuracy in mapping the crop types, which can be further utilized for effective management of the agricultural sectors.