Fine classification of crops in Karamay region based on time series

Classification principle

Time-series-based fine crop classification methods analyze significant changes in crop spectral characteristics, vegetation indices (such as NDVI), and moisture indices (such as NDWI) throughout the entire growth cycle, and combine this with advanced classification models to achieve refined classification of crop types. This method is effectively adaptable to agricultural scenarios with complex crop types and wide geographical areas, and has significant application value, especially in crop planting area statistics and regional agricultural monitoring.

Classification steps

1. Extraction of land parcel boundaries 

Remote sensing images of the study area were acquired by selecting long-term remote sensing image data (long-term synthetic images processed with cloud cover). During image segmentation, the large-scale mean-shift (LSMS) image segmentation method was used to extract the boundaries of crop plots within the study area, generating vector boundary data for each plot. This plot boundary data forms the basis for subsequent classification, is suitable for object-oriented classification techniques, and helps to accurately distinguish adjacent crop plots.

2. Extraction of crop phenological characteristic curves

Using time-series satellite remote sensing images, vegetation index (NDVI) and moisture index (NDWI) of plots at different time points are extracted and mapped to generate time-series characteristic curves for each plot.

These characteristic curves not only visually reflect crop growth patterns but also help infer key phenological time points, such as sowing, rapid growth, maturity, and withering stages. Based on the analysis of these phenological time points, suitable time windows for crop classification can be determined, thereby optimizing classification parameters and improving classification accuracy.

3. Classification feature establishment

Based on plot vector data, feature values ​​for classification windows are extracted from multidimensional raster data using a zonal statistical method. Vegetation indices, phenological parameters, and other auxiliary features within different time windows together constitute the feature dataset required for classification.

This feature-building method based on time-series statistics can effectively capture the spatial and temporal differences of crops, providing high-quality input data for object-oriented crop classification.

4. Model training

Using user-provided ground validation data as labels, a training dataset was built to train a machine learning classification model. Within the study area, the main crop types were alfalfa, maize, cotton, and other crops. The model training process involved cross-validation and hyperparameter tuning to ensure the generalization ability and accuracy of the classification model.

5. Model application

The trained classification model is then extended to a larger area, and crop classification is performed using time-series remote sensing data. During this process, the planting area of ​​different crops is calculated based on the classification results.

Based on the classification results, further work can be carried out on regional crop distribution analysis, agricultural production statistics and dynamic monitoring, etc., to provide a scientific basis for agricultural management.

Summarize

Classification methods based on time-series features can adapt to diverse crop planting areas and complex planting patterns, achieving high classification accuracy even with complex crop types. Because this method utilizes rich time-series feature information, it reduces reliance on ground-labeled samples, making it suitable for crop classification tasks over large areas. By fully leveraging the spatial coverage advantages of satellite remote sensing data, it can achieve large-scale crop distribution and area statistics, supporting regional and national agricultural planning.

Figure 1. Schematic diagram of time series characteristic curves for different crop types

Figure 2. Results of plot boundary extraction

Figure 3. Object-oriented classification results