A Multi-Module Spatial Semantic Network with Gradient Boosting for Efficient Crop Classification

For the past few years, accurate crop classification is critical for efficient agricultural management and monitoring. While traditional classification techniques have leveraged spectral indices, they often struggle to fully capture the complex and diverse multispectral features present in agricultural landscapes. This challenge is particularly evident when differentiating crops with similar spectral signatures, leading to misclassification and reduced accuracy in traditional models. To address this problem, a hybrid approach that combines a deep learning model namely Multi-module Spatial Semantic Network (MSSNet) with the Gradient Boosting Classifier (GBC) is proposed. The dataset utilized in this research is sentinel-2, where the median filter is used to remove noise at the preprocessing stage. Then, MSSNet architecture is applied in feature extraction where pre-trained residual network (ResNet50), embedded receptive field module (RFB) and skip connection are incorporated to extract multi-spectral and spatial features. Finally, the extracted features are fed into a GBC, which is suitable for the robust performance in pixel mapping. This MSSNet+GBC combination leverages the strengths and offers improved accuracy in distinguishing between crop types and conditions. The experimental results demonstrate that MSSNet+GBC outperforms conventional methods such as 3D-Convolutional Neural Network (CNN) with higher accuracy of 98.67% using a sentinel-2 dataset for precise crop classification.