Long-term yield assessment based on time-series remote sensing images

This project, based on time-series satellite remote sensing imagery and combined with plot delineation, phenological period analysis, and historical yield performance evaluation, conducted a long-term yield assessment study for target plots. Through precise data processing and analysis, it provides a scientific basis for plot yield performance and contributes to precision agricultural management.

1. Land parcel boundary delineation 

Based on the provided reference plot (Sanfenchang), some other plots were divided, totaling 40, for subsequent analysis.

The overall distribution of land parcels is as follows, with the numbers in the parcels representing parcel IDs.

Figure 1. Site Schematic Diagram

2. Phenological Period Determination 

Using time-series satellite remote sensing imagery, the NDVI vegetation index of all plots from 2023 to 2024 was mapped. The periodic changes in the vegetation index were used to infer the planting season and the specific phenological period, providing a reference window for subsequent yield forecasting.

Figure 2. Time series of NDVI vegetation index for all plots from 2023 to 2024

The above figure shows the relative position of the target plot to all plots in terms of growth. The overall growth throughout the year is close to the average, indicating that the growth of the target plot is at the average level in a wide range. This is also reflected in the subsequent long-term yield performance assessment chart and the annual yield distribution chart.

Figure 3. Target Plots and All Plots in 2023-2024  Time series plot of average NDVI vegetation index.

Figure 4. Box plot of average NDVI vegetation index for the target plot and all plots in 2023-2024

Figures 3 and 4 show the comparison results between the target plot and the average level. Type 1 represents the target plot, and type 2 represents the average of the plots. It can be seen that during the seedling stage before winter, the growth of the target plot was worse than the average. Although the growth during the subsequent jointing stage was the same as or slightly higher than the average, due to the tillering characteristics of gramineous crops, the effective tillering level before winter was lower than the average, so the final yield was at the average level.

This step used data from all plots in the 2023-2024 growing season and performed dynamic time warping and linear interpolation to generate continuous time-series images for accurate analysis.

3. Long-cycle production performance assessment 

Figure 5. Long-cycle production performance evaluation chart

The chart above shows the statistical information on yield performance over the past five years. Long-term yield performance information can be extracted from the cyclical yield fluctuations of crops to more comprehensively and accurately assess the yield performance of a plot, identify plot problems, and make targeted improvements.

The numbers in the plot represent the net number of years with higher yields for that plot over the past five years. The value range is -5 to 5. The calculation method is as follows: Annual yield performance is statistically analyzed and clustered using a clustering algorithm. Each year's yield performance is categorized as good or poor. Good means that the yield that year was above average among all plots, while poor means below average. The yield performance for each year over the past five years is calculated, and then the number of years with good yield performance is subtracted from the number of years with poor yield performance. This gives the numbers in the plots above. Positive values ​​indicate that the number of years with good yield performance is greater than the number of years with poor yield performance, indicating an overall yield preference. Negative values, on the other hand, indicate the opposite.

Figure 6. Distribution of Production of All Plots in 2024

Figure 6 shows the 2024 yield of each plot, all per mu (unit of land area). Target plots are marked with white borders. Red and blue legends represent yields that are lower/higher than the average, respectively. One-dimensional unsupervised clustering was used for classification, which, compared to simple average comparison, can better handle outliers and boundary cases, increasing the reliability of the classification. The map data is the same as the yield statistics table data.

Overall, the difference between the lowest and highest yield plots is 100-170 jin/mu. For each plot, the yield varies from year to year, ranging from 10-50 jin/mu, and the annual yield differences between different plots also vary.

Through the above production analysis, we can gain a deeper understanding of the annual production distribution, identify plots with high production potential, and carry out targeted management.