Object-oriented extraction of wheat plots based on hyperspectral remote sensing

In precision agriculture practices, plot boundaries are the most fundamental planting data. They provide a link between all crop data and the real world, provide accurate geographical boundaries for plot-level growth statistics and analysis, and also provide a foundation for analyzing plot area changes and long-term monitoring of crops at various phenological stages.

Due to the different agricultural production models and planting history in different regions, the definition and concept of "plot" also vary, and the results of observations at different scales are not entirely consistent. The plot extraction in this paper is based on 10-meter resolution satellite remote sensing images. Therefore, a plot is defined as a relatively independent planting unit divided by paved roads such as rural roads or village roads.

Wheat Plot Image Extraction Method

The CAS Xiguang Aerospace team used the maximum value composite image of wheat in Lintong District during the jointing stage in 2023, and extracted wheat plots for the 2023-2024 season using object-oriented classification technology.
Object-oriented classification divides land features into independent objects through a segmentation-then-classification approach, and then classifies them based on these objects. This approach largely avoids the salt-and-pepper effect of traditional pixel-based classification while increasing the interpretability of the classification. The output is a land parcel vector, which can be directly used for shape and area calculations and statistics, facilitating subsequent analysis.

▲Figure 1 Object-oriented classification results

Wheat Image Results Analysis

Based on literature and local landform characteristics, the categories were divided into five types: wheat, urban areas/man-made structures, bare soil, sandy areas/highly reflective objects, and water bodies. The classification results for water bodies and urban areas were relatively satisfactory. Some roads were confused with bare soil, and some rooftops had high reflectivity, affecting the classification results. Therefore, merging highly reflective objects with sandy areas increased the reliability of the classification results.

It can visually display the urban area and unused land. It can also calculate the area of ​​corresponding land features year by year and analyze the development and changing trends of towns and various land use types.

▲Figure 2 Accuracy evaluation of object-oriented random forest classifier

Xi'an Institute of Optics and Precision Mechanics CAS Aerospace Science and Technology Group Co.,LTD - Agricultural Plot Growth Analysis

In this case, a total of 409 samples from five land cover categories were visually interpreted and labeled. 80% were used for training and 20% for validation. Figure 2 shows the classifier's accuracy evaluation, which demonstrated good classification performance for farmland, bare soil, and water bodies. The classification accuracy for urban areas/man-made structures was relatively poor, possibly due to the complex distribution of land cover and limitations imposed by satellite imagery resolution.

▲Figure 3. Partial wheat extraction in Xiangqiao Subdistrict, Lintong District

The image above shows the extraction results for a local area. The overall effect is good, with few missed or incorrect extractions. However, farmland near main roads and urban areas is somewhat confused with roads. Some plots are quite small and are mixed with farm roads and village roads.

▲Figure 4 Wheat extraction in the mountainous area of ​​Lintong

The figure shows the extraction results of wheat plots in the southern mountainous area of ​​Lintong District using the same classification model. The results are good, indicating that image segmentation can effectively obtain texture information and has good extraction results for wheat plots with different terrains and shapes.

▲Figure 5 Distribution map of wheat planting area in various streets of Lintong District

The image shows the distribution of wheat planting area in various streets of Lintong District. The darker the color, the larger the wheat planting area.

It is evident that most wheat fields are concentrated in the northern part of Lintong, with Xiangqiao Subdistrict having the largest planting area at 66,602 mu. The southern area is mountainous, with a small amount of wheat planted there.

▲ Table 1. Statistics on Planting Area in Various Subdistricts of Lintong District

▲Figure 6 Distribution map of wheat plots in various subdistricts of Lintong District

The image shows the distribution of wheat planting area in various streets of Lintong District. The darker the color, the more wheat plots there are.

It is evident that most wheat fields are concentrated in the northern part of Lintong, with the largest number (1950) in Xiangqiao Subdistrict. The southern area is mountainous, with a small amount of wheat cultivation.

▲ Table 2 Statistical Table of Wheat Plots in Each Subdistrict of Lintong District

▲Figure 7 Distribution of wheat growth anomalies during the jointing stage in various subdistricts of Lintong District

Based on the extraction of land parcels, the differences between the annual growth and the growth in recent years can be analyzed for each land parcel within a specified time range.

As shown in the figure, the darker the red, the worse the crop is compared to the past five years; the darker the green, the better the crop is compared to the past five years.

Anomaly analysis can analyze the performance of each plot of land each year, identify plots with stable and good yields and those with large yield fluctuations and poor performance, and help managers to make large-scale planting plans.