Early-Season Crop Classification Utilizing Time Series Based Deep Learning with Multi-Sensor Remote Sensing Data

Early-season crop type classification provides crucial information for monitoring in-season crop growth and predicting crop diseases, supporting global food security. In remote sensing practices for agriculture, it is technically challenging to perform early-season prediction mainly due to factors such as significant soil interference to crops on the ground and limited availability of remote sensing data. In this work, we propose a time series based machine/deep learning approach which is able to utilize all useful features of satellite images captured by multiple sensors (RCM, Sentinel-1, and Sentinel-2) in the early season. The algorithm is implemented iteratively on data sequences such that when a new satellite image becomes available, new features will be appended to existing time series data, thus providing a more accurate prediction map in updated regions. In the study, four (4) machine/deep learning algorithms, including XGBoost, CNN, and LSTM, are evaluated for their performance in terms of prediction accuracy and computational complexity. The proposed approach is tested in three study areas in the Canadian Prairie provinces in the 2021 season. Crop type mapping results are obtained from the available early-season images from June 10 to June 30, 2021, with extended results to July 31, 2021. The validation results on independent crop survey data show that the proposed algorithms can achieve crop type classification accuracy around 85% at the end of June and above 90% at the end of July.