Early- and in-season crop type mapping without current-year ground truth: Generating labels from historical information via a topology-based approach

Land cover classification in remote sensing is often faced with the challenge of limited ground truth labels. Incorporating historical information has potential to significantly lower expensive cost associated collecting and, more importantly, enable early- and in-season mapping that helpful many pre-harvest decisions. In this study, we propose a new approach can effectively transfer knowledge about topology (i.e. relative position) different crop types spectral feature space (e.g. histogram SWIR1 vs RDEG1 bands) generate labels, thereby supporting year. Importantly, our does not attempt decision boundaries are susceptible inter-annual variations weather management, but relies on robust shift-invariant information. We tested for corn/soybeans US Midwest, paddy rice/corn/soybeans Northeast China multiple crops Northern France using Landsat-8 Sentinel-2 data. Results show automatically generates high-quality labels target year immediately after each image becomes available. Based these generated from approach, subsequent type random forest classifier reach F1 score as high 0.887 corn early silking stage 0.851 soybean flowering overall accuracy 0.873 test state Iowa. China, scores rice, soybeans exceed 0.85 two half months ahead harvest. Hauts-de-France region, OA could 0.837 based approach. Overall, results highlight unique advantages transferring maximizing timeliness maps. Our supports general paradigm shift towards learning transferrable generalizable facilitate land classification.