EMET: An emergence-based thermal phenological framework for near real-time crop type mapping

Abstract

Near real-time (NRT) crop type mapping plays a crucial role in modeling crop development, managing food supply chains, and supporting sustainable agriculture. The low-latency updates on crop type distribution also help assess the impacts of weather extremes and climate change on agricultural production in a timely fashion, aiding in identification of early risks in food insecurity as well as rapid assessments of the damage. Yet NRT crop type mapping is challenging due to the obstacle in acquiring timely crop type reference labels during the current season for crop mapping model building. Meanwhile, the crop mapping models constructed with historical crop type labels and corresponding satellite imagery may not be applicable to the current season in NRT due to spatiotemporal variability of crop phenology. The difficulty in characterizing crop phenology in NRT remains a significant hurdle in NRT crop type mapping. To tackle these issues, a novel emergence-based thermal phenological framework (EMET) is proposed in this study for field-level NRT crop type mapping. The EMET framework comprises three key components: hybrid deep learning spatiotemporal image fusion, NRT thermal-based crop phenology normalization, and NRT crop type characterization. The hybrid fusion model integrates super-resolution convolutional neural network (SRCNN) and long short-term memory (LSTM) to generate daily satellite observations with a high spatial resolution in NRT. The NRT thermal-based crop phenology normalization innovatively synthesizes within-season crop emergence (WISE) model and thermal time accumulation throughout the growing season, to timely normalize crop phenological progress derived from temporally dense fusion imagery. The NRT normalized fusion time series are then fed into an advanced deep learning classifier, the self-attention based LSTM (SAtLSTM) model, to identify crop types. Results in Illinois and Minnesota of the U.S. Corn Belt suggest that the EMET framework significantly enhances the model scalability with crop phenology normalized in NRT for timely crop mapping. A consistently higher overall accuracy is yielded by the EMET framework throughout the growing season compared to the calendar-based and WISE-based benchmark scenarios. When transferred to different study sites and testing years, EMET maintains an advantage of over 5% in overall accuracy during early- to mid-season. Moreover, EMET reaches an overall accuracy of 85% a month earlier than the benchmarks, and it can accurately characterize crop types with an overall accuracy of 90% as early as in late July. F1 scores for both corn and soybeans also achieve 90% around late July. The EMET framework paves the way for large-scale satellite-based NRT crop type mapping at the field level, which can largely help reduce food market volatility to enhance food security, as well as benefit a variety of agricultural applications to optimize crop management towards more sustainable agricultural production.