Large Area Crops Mapping by Phenological Horizon Attention Transformer (PHAT) Method Using MODIS Time-Series Imagery

Abstract

Accurate collection of crop planting information at large area is essential for estimating agricultural productivity and ensuring food security. Different crops have varying growth cycles and phenological stages, and changes in factors such as topography, soil type, and moisture conditions can lead to diversity in crops growth status, which complicates uniform monitoring. Multiple crops mapping simultaneously with high precision presents a significant challenge due to the high spatial heterogeneity of crops distribution across vast regions. To address these challenges, this article developed an advanced deep learning crop mapping method, i.e., phenological horizon attention mechanism-transformer model (PHAT) to achieve rapid and accurate multiple crops extraction over large areas. Initially, time-series data were constructed using the normalized differential vegetation index (NDVI) dataset based on moderate resolution imaging spectroradiometer (MODIS) product. Subsequently, in the mixed pixel decomposition phase, orthogonal subspace projection and vertex component analysis were employed to identify crop types and extract endmembers. While the regular changes in the time-series NDVI reflect the phenological evolution trend among multiple crops, but the phenological characteristics difference between the same crop is extremely difficult to find. The PHAT model was therefore trained using the phenological features of endmembers to obtain the spatial distribution of crops, and to resolve the issue of varying time-series curves for the same crop across large areas. This study selected the North China Plain in 2021 as the research area, utilizing Google Earth data and Landsat 8 images to verify the approach's accuracy. Based on the MODIS NDVI data with a coarse spatial resolution of 250 m, our method achieved an OA of 90.1% for the synchronous extraction of soybean, spring peanut-summer sesame, winter wheat-summer maize, paddy rice, and rapeseed-cotton, with a RMSE of approximately 12% in 16.6 million mu of cultivated land.