Automatic rice fields mapping by Fusion of in-Decoder CNN and Data Augmentation Techniques on Landsat-8 multi-temporal images

Rice is the main food for the world's people. Monitoring and mapping rice fields play an important role in agricultural planning. Nowadays, intelligent management of rice fields has improved by remote sensing technology and deep learning algorithms. This research aims to study is the Fusion in-Decoder model and Data Augmentation techniques by using extracted multi-temporal maps of NDVI, LST, and LSWI indices from Landsat-8 images for mapping rice fields at the state of California, in 2020. Therefore, six architectures of Fusion in-Decoder model were designed, after radiometric corrections, atmospheric corrections, and generate multi-temporal maps of NDVI, LST, and LSWI indices, and simulation of different phenologies of rice crop with the shift of multi-temporal indices and PCA algorithm: (1) One Encoder-one Decoder (NDVI) and use of Data Augmentation techniques by the shift of multi-temporal indices and PCA algorithm, (2) Two Encoders-one Decoder (NDVI-LST) and use of Data Augmentation techniques by the shift of multi-temporal indices and PCA algorithm, (3) Two Encoders-one Decoder (NDVI-LSWI) and use of Data Augmentation techniques by the shift of multi-temporal indices and PCA algorithm, (4) Three Encoders-Decoder (NDVI-LST-LSWI) and use of Data Augmentation techniques by the shift of multi-temporal indices and PCA algorithm, (5) Three Encoders-one Decoder (NDVI-LST- LSWI) and use of Data Augmentation technique by the shift of multi-temporal indices, and (6) Three Encoders-one Decoder (NDVI-LST-LSWI) without the use of Data Augmentation techniques. The fusion in-decoder and Data Augmentation techniques compared with four classifiers Decision Tree (DT), Logistic Regression (LR), Multi-Layer Perceptron (MLP), and Auto-Encoder (AE). The results showed that the Fusion in-Decoder model with three Encoders-one Decoder (NDVI-LST-LSWI) and use of Data Augmentation techniques by the shift of multi-temporal indices and PCA algorithm performed best with Kappa coefficient (89/85%) for multi-temporal images of months April to August at the state of California. Besides, among the comparison classifiers, AE showed the worst result with Kappa coefficient (31.88%).