Cross-Year Reuse of Historical Samples for Crop Mapping Based on Environmental Similarity

Abstract

Crop classification maps are fundamental data for global change research, regional agricultural regulation, fine production, and insurance services. The key to crop classification is samples, but it is very time-consuming in annual field sampling. Therefore, how to use historical samples in crop classification for future years at a lower cost is a research hotspot. By constructing the spectral feature vector of each historical sample in the historical year and its neighboring pixels in the target year, we produced new samples and classified them in the target year. Specifically, based on environmental similarity, we first calculated the similarities of every two pixels between each historical year and target year and took neighboring pixels with the highest local similarity as potential samples. Then, cluster analysis was performed on those potential samples of the same crop, and the class with more pixels is selected as newly generated samples for classification of the target year. The experiment in Heilongjiang province, China showed that this method can generate new samples with the uniform spatial distribution and that the proportion of various crops is consistent with field data in historical years. The overall accuracy of the target year by the newly generated sample and the real sample is 61.57 and 80.58%, respectively. The spatial pattern of maps obtained by two models is basically the same, and the classification based on the newly generated samples identified rice better. For areas with majority fields having no rotation, this method overcomes the problem of insufficient samples caused by difficulties in visual interpretation and high cost on field sampling, effectively improves the utilization rate of historical samples, and provides a new idea for crop mapping in areas lacking field samples of the target year.

Cropformer: A new generalized deep learning classification approach for multi-scenario crop classification

“…These models are sensitive to the spectral information of crops and use vegetation indices and spectral values as the main feature inputs. However, the sequence relationships hidden in the time series are not exploited, so more temporal features are incorporated in the model inputs including the statistical value of spectral value and vegetation indices and statistical features of vegetation indices curve (Pelletier et al, 2016;Zhang et al, 2018;Zeng et al, 2020;Liu et al, 2021). Comparing multiple crop vegetation index curves and obtaining key dates and key observations from the curves to distinguish crops can be effective in improving classification performance (Simonneaux et al, 2008;Lebourgeois et al, 2017).…”

Crop Mapping without Labels: Investigating Temporal and Spatial Transferability of Crop Classification Models Using a 5-Year Sentinel-2 Series and Machine Learning

“…Although site conditions and environmental factors may be better suited for explaining natural vegetation development in areas without abrupt agronomic interventions, their inclusion in crop classification models can still provide valuable insights. For instance, some studies have examined the role of environmental similarity as a factor for label generation under unseen conditions [40]. Indeed, meaningful environmental variables like digital elevation models (DEM) can shed light on altitudinal phenology variations.…”

Mapping Main Grain Crops and Change Analysis in the West Liaohe River Basin with Limited Samples Based on Google Earth Engine

“…However, these methods above rely on the quantity of samples to train a robust classification model, particularly for deep learning algorithms [28]. Sample collection is currently dominated by field investigation and manual interpretation via high-resolution images, which is time-consuming and labor-intensive, making it difficult to collect data evenly in large study areas and to ensure that data are available over multiple years [29]. Several studies attempted to perform classification for target years with a lack of sample data by exploiting archived samples, involving sample augmentation and feature or model migration [28,30,31].…”

“…Zhang et al [32] proposed a method that utilized samples from 2013 to 2017 to determine unchanged pixels as new samples for 2018. Liu et al [29] calculated spectral similarity between the historical year and target year of samples and performed cluster analysis to produce new samples. Transferring models trained on archived reference data to identify crops in the target year is also an approach to solving this issue.…”