Abstract:Hyperspectral remote sensing images, due to their unique characteristics, can provide rich spectral features and spatial information. Crops have different absorption and reflection characteristics in different bands, which gives hyperspectral images a natural advantage in crop classification. The academic community has also made progress in this area through a series of studies. This is of great significance for accurately obtaining the spatial distribution information of crops and realizing precision and intelligent management in agricultural production. However, hyperspectral data have low spatial resolution and suffers from mixed pixel interference and lack of spatial information description for crops. Meanwhile, influenced by the phenological characteristics of crops, different crops respond differently to the same spectrum during the same phenological period, and similarly, the spectral response of the same crop varies at different phenological stages. This phenomenon leads to typical intra-class and inter-class feature differences, which increase the complexity of hyperspectral crop classification. To address the issues of missing spatial information and inconsistent spectral responses, We propose a prior-guided and data-driven deep learning method for crop classification in hyperspectral remote sensing imagery, named Cross-scale Enhancement Network (CSENet). This method couples index features and incorporates them as prior knowledge to the network. Firstly, we utilize the phenological differences of different crops to select the spectral bands with significant responses and perform feature dimension reduction; Secondly, we utilize the high reflectance characteristics of vegetation in the near-infrared band and introduce multiple classic vegetation indices to construct a spectral feature library as prior knowledge to enrich the network features. Finally, a cross-scale enhancement attention mechanism is designed to obtain multi-scale crop features and compensate for the lack of spatial description information of crops. In addition, a cross-channel enhancement module is further constructed to deeply explore the spectral information and solve the problem of confusion caused by similar spectral characteristics, making the network more robust and resistant to interference in crop classification task of hyperspectral imagery. Unlike the traditional pixel-by-pixel approach for hyperspectral classification, CSENet employs a fully convolutional network architecture with a markedly reduced parameter count and faster inference speed. We selected typical crops such as wheat, rice, rape, and potatoes to evalute algorithm in Inner Mongolia. The experimental results demonstrate that the proposed method has a good feature classification ability for typical crops and also has the conditions for practical application in terms of time efficiency.

A Neural Network Method for Classification of Sunlit and Shaded Components of Wheat Canopies in the Field Using High-Resolution Hyperspectral Imagery

Deep Learning-Based Crop Mapping in the Cloudy Season Using One-Shot Hyperspectral Satellite Imagery.

Classifying Wheat Hyperspectral Pixels of Healthy Heads and Fusarium Head Blight Disease Using a Deep Neural Network in the Wild Field

Mix Convolutional Neural Networks for Hyperspectral Wheat Variety Discrimination

Modeling the spatial-spectral characteristics of plants for nutrient status identification using hyperspectral data and deep learning methods

Identification of Defective Maize Seeds Using Hyperspectral Imaging Combined with Deep Learning

In-field chlorophyll estimation based on hyperspectral images segmentation and pixel-wise spectra clustering of wheat canopy

Hyperspectral Imaging for Phenotyping Plant Drought Stress and Nitrogen Interactions Using Multivariate Modeling and Machine Learning Techniques in Wheat

Plant Species Classification Based on Hyperspectral Imaging via a Lightweight Convolutional Neural Network Model

Novel 3-D Deep Neural Network Architecture for Crop Classification Using Remote Sensing-Based Hyperspectral Images

Nondestructive Identification of Wheat Species using Deep Convolutional Networks with Oversampling Strategies on Near-Infrared Hyperspectral Imagery

Hyperspectral imaging with shallow convolutional neural networks (SCNN) predicts the early herbicide stress in wheat cultivars

Wheat Kernel Variety Identification Based on a Large Near-Infrared Spectral Dataset and a Novel Deep Learning-Based Feature Selection Method

Novel 3D Deep Neural Network Architecture for Crop Classification Using Remote Sensing-Based Hyperspectral Images

Identification of Slightly Sprouted Wheat Kernels Using Hyperspectral Imaging Technology and Different Deep Convolutional Neural Networks

Reflectance Images of Effective Wavelengths from Hyperspectral Imaging for Identification of Fusarium Head Blight-Infected Wheat Kernels Combined with a Residual Attention Convolution Neural Network

Determination of Wheat Heading Stage Using Convolutional Neural Networks on Multispectral UAV Imaging Data

New Methodology of Hyperspectral Information Extraction and Accuracy Assessment Based on a Neural Network.

Nondestructive Detection of Sunflower Seed Vigor and Moisture Content Based on Hyperspectral Imaging and Chemometrics

Crop Type Classification Network for Hyperspectral Images Coupled with Agricultural Vegetation Indices

Using VIS-NIR hyperspectral imaging and deep learning for non-destructive high-throughput quantification and visualization of nutrients in wheat grains