Integration of high-dimensional remote sensing feature domains: A multi-branch MLP with reconfigurable input for advanced crop classification

Timely and accurate crop distribution mapping is critical for agricultural production, yet large-scale crop type identification using remote sensing remains challenging. Conventional classification models, constrained by single and rigid data structures, often fail to fully utilize the information within multimodal temporal remote sensing feature spaces, limiting accuracy improvements. This study proposes a novel Multi-branch multilayer perceptron (M-MLP) model that addresses this limitation by dynamically reconfiguring input data organization to mine multiple feature domains. The model leverages complementary information through decision-level fusion, simultaneously incorporating phenological patterns from time series and discriminative features from multimodal data to fully exploit high-dimensional feature spaces. We constructed a large-scale crop classification dataset containing 561,651 samples derived from Sentinel-1/2 time series imagery for evaluation. Comparative analysis against six state-of-the-art models demonstrated M-MLP's superior performance, achieving optimal results of 97.31 % Overall Accuracy (OA) and 96.58 % Macro-F1—representing improvements of 1-6 % over existing approaches. The integrated four-branch configuration significantly enhanced accuracy (1.2 % OA and Macro-F1 gains), with each branch contributing positively to performance. This research elucidates the distinct roles of temporal versus spatial-spectral domains in crop classification and demonstrates M-MLP's early-season classification capability. The framework establishes "high-dimensional remote sensing feature domain integration" as a generalized modeling paradigm, offering new avenues for advancing crop classification methodologies.