Automatic crop type mapping based on crop-wise indicative features

Abstract

Mapping crop types with remote sensing imagery is crucial for timely acquisition of crop planting information. However, large-scale crop type mapping is often hindered by the absence of ground truth data and spatiotemporal heterogeneity of available cloud-free optical images. In this study, the conception of daily crop-wise indicative features (DCIF) was proposed for automatic crop mapping to address this issue. Crop-wise indicative feature (CIF) was defined as a comprehensive feature that can distinctly differentiates the target crop and all other classes, with the CIF value showing a polarization trend with target crop close to 1 and others close to 0. Logistic Regression (LR) model was utilized to learn CIF of target crop on each single date, referred to as CIF extractor. The time series analysis method (i.e., seasonal trend decomposition) and imputation method was then utilized for the discrete time series of CIF extractors, which were separately trained based on training samples on each day. This process yielded the DCIF extractor that can capture the unique feature pattern on any given day during the entire growing period. The comprehensive features produced by DCIF extractor later served as the input of Otsu algorithm to automatic distinguish target crop from other crops. Our results in Heilongjiang, China, Iowa, USA, and Bas-Rhin, France show that the DCIF extractor effectively overcomes abovementioned challenges in crop mapping by applying time series analysis method at model-level. The overall accuracy exceeded 90 % in Heilongjiang and Iowa for both 2021 and 2022. In Bas-Rhin, where fragmented land parcels and uneven crop distribution are prevalent, the overall accuracy surpassed 87 %. The proposed method provides a scientific basis for crop acreage estimation and field management, contributing to more informed agricultural practices.