Our research focused on the critical challenge of accurately extracting canopy temperature from cotton crops during the flowering and boll-setting stage using drone technology. The primary objective was to determine how variations in flight altitude and different methods for removing outlier temperature data influence the precision of the final extraction. We believe that refining these parameters is essential for advancing drone technology in precision agriculture, particularly for monitoring plant water stress and optimizing irrigation scheduling in water-scarce regions like drip-irrigated cotton fields.
To achieve this, we conducted a two-year field experiment, gathering data during the 2023 and 2024 growing seasons. We utilized a commercial drone equipped with a thermal infrared camera. Our experimental design involved capturing thermal images of cotton canopies at five distinct flight altitudes: 12 m, 20 m, 30 m, 50 m, and 70 m. The cotton was subjected to four different irrigation frequencies (3, 5, 7, and 12 days per irrigation event) to create a range of canopy temperature conditions. For each flight altitude, we systematically compared three distinct strategies for processing the raw thermal data to extract the pure canopy temperature signal, aiming to minimize the influence of non-canopy elements like soil and sensor noise. These strategies included the top and bottom 1% elimination method, the low-frequency 0.5% elimination method, and the low-frequency 1% elimination method.
Our fundamental analysis assessed the correlation between the drone-extracted canopy temperatures and the ground-truth temperatures measured with a handheld infrared thermometer. The relationship between the extracted canopy temperature (Textracted) and the measured canopy temperature (Tmeasured) was analyzed using a simple linear regression model:
$$ T_{extracted} = a \cdot T_{measured} + b $$
where a is the slope of the regression line and b is the intercept. The accuracy of this relationship was quantified using the coefficient of determination (R²) and the root mean square error (RMSE). The RMSE was defined as:
$$ RMSE = \sqrt{ \frac{1}{n} \sum_{i=1}^{n} (T_{measured, i} – T_{extracted, i})^2 } $$
where n represents the number of sample points.
Our analysis of the data from both years revealed that the optimal method for outlier removal was highly dependent on the flight altitude. For lower flight altitudes, specifically the 12 to 30 meter range, we found that the low-frequency 0.5% elimination method yielded the most accurate results. This method involves removing the lowest 0.5% of temperature values from the canopy temperature histogram, which are often caused by shaded soil or mixed pixels at the edges of the canopy. For higher flight altitudes of 50 and 70 meters, the top and bottom 1% elimination method proved to be superior.
The following table summarizes the performance of the preferred outlier removal methods for each altitude range across both years of our study.
| Altitude Range | Preferred Method | Year | Model Fit (R²) | RMSE (°C) |
|---|---|---|---|---|
| Low Altitude (12 – 30 m) | Low-Frequency 0.5% | 2023 | 0.874 | 2.435 |
| 2024 | 0.934 | 2.171 | ||
| High Altitude (50 – 70 m) | Top & Bottom 1% | 2023 | 0.833 | 3.904 |
| 2024 | 0.914 | 3.859 |
The data in Table 1 clearly demonstrates the advantage of using a tailored outlier removal strategy for different altitudes. The low-frequency 0.5% method for low-altitude drone flights consistently provided a more accurate prediction of ground-measured temperatures, as evidenced by the higher R² and lower RMSE values. The superiority of this method for low-altitude drone technology likely stems from its ability to precisely filter out the temperature of the cooler soil background, which becomes more mixed with canopy pixels at these closer ranges. Conversely, at higher altitudes, the larger footprint of each pixel leads to more temperature averaging, making the symmetrical removal of extreme values from both ends of the histogram the most effective approach for noise reduction.
After identifying the best method for each altitude range, our next step was to compare the performance of the five different flight altitudes (12, 20, 30, 50, and 70 m) using their respective optimal outlier removal strategies. The goal was to pinpoint the single best flight altitude for deploying drone technology to monitor cotton canopy temperature.
The comparative analysis of different flight altitudes showed a clear trend. The models built using data from lower altitudes (12 to 30 m) consistently outperformed those from higher altitudes (50 to 70 m). Across both years, the 30-meter flight altitude emerged as the most robust choice, offering the best balance between accuracy and operational practicality. The following table presents a comprehensive summary of the model performance for each altitude over the two years of our study.
| Flight Altitude (m) | 2023 | 2024 | ||||||
|---|---|---|---|---|---|---|---|---|
| Fitting Equation | R² (Train) | RMSE (Train) | R² (Validation) | Fitting Equation | R² (Train) | RMSE (Train) | R² (Validation) | |
| 12 | y = 0.919x – 0.408 | 0.709 | 3.136 | 0.855 | y = 0.795x + 4.093 | 0.740 | 2.313 | 0.829 |
| 20 | y = 1.025x – 3.331 | 0.819 | 2.614 | 0.830 | y = 0.752x + 5.346 | 0.783 | 2.345 | 0.917 |
| 30 | y = 1.607x – 21.630 | 0.863 | 2.424 | 0.923 | y = 1.014x – 3.949 | 0.720 | 3.664 | 0.766 |
| 50 | y = 0.893x – 0.118 | 0.677 | 3.653 | 0.810 | y = 0.774x + 3.243 | 0.701 | 3.804 | 0.817 |
| 70 | y = 0.930x – 1.980 | 0.741 | 4.285 | 0.541 | y = 0.788x – 1.690 | 0.687 | 4.904 | 0.737 |
As shown in Table 2, the 30 m drone flight altitude consistently achieved the highest training R² (0.863 in 2023 and 0.720 in 2024), indicating a strong linear relationship between the drone-extracted and ground-measured temperatures. Although the performance of the 30 m altitude fluctuated slightly between the two years, it remained the most stable and predictive overall, particularly when considering the model’s validation performance. The validation R² for the 30 m altitude reached 0.923 in 2023, the highest value recorded across all altitudes and years. In 2024, while the model at 20 m showed a slightly higher validation R² of 0.917, the 30 m altitude demonstrated superior consistency in its training and validation results without drastic overfitting.
The inferior performance at 12 m and 20 m altitudes, while offering higher spatial resolution, often resulted in more complex images with greater canopy shading and soil background noise, which challenged the simple outlier removal methods. At 50 m and 70 m, the reduced spatial resolution led to a higher incidence of mixed pixels, where each pixel represents a blend of canopy, soil, and shadow. This blending effect dampened the thermal signal, making it more difficult to accurately isolate the pure canopy temperature and leading to higher RMSE values, which were notably around 3.9°C and 3.86°C for the high-altitude optimized method.
Further analysis of the statistical characteristics of the extracted temperatures confirmed these findings. Processing the data with the optimal outlier removal methods for each altitude range consistently reduced the variance in the thermal data. For instance, when using the low-frequency 0.5% method for the 12-30 m range, the mean extracted temperature shifted closer to the mean measured temperature, and the coefficient of variation (CV) increased slightly, reflecting the removal of cold, low-variance soil pixels and a greater emphasis on the more dynamic canopy signal. The standard deviation of the extracted temperatures was also reduced after filtering, indicating a more stable and reliable dataset for subsequent analysis.
The practical implications of this research for drone technology in agriculture are significant. By establishing that a 30 m flight altitude, combined with a low-frequency 0.5% outlier removal method, provides the most accurate and reliable canopy temperature data for drip-irrigated cotton, we have provided a clear and actionable guideline for field practitioners and researchers. This specific configuration of drone technology can dramatically improve the accuracy of crop water stress index (CWSI) calculations and other thermal-based diagnostics. In our study, the prediction accuracy of the canopy temperature model improved by 70.71% in 2023 and 4.03% in 2024 when applying this optimal configuration compared to non-optimized approaches, highlighting the substantial gains achievable through careful parameter selection.
In conclusion, our intensive two-year study systematically demonstrates that the precision of canopy temperature extraction using drone technology is not solely a function of the hardware but is profoundly influenced by both the flight altitude and the data processing methodology. For optimal results in cotton fields, we strongly recommend flying at a height of 30 meters and applying a low-frequency 0.5% elimination method to the thermal data. This combination effectively balances the need for high-resolution imagery to distinguish canopy from soil with the need for a robust, scalable data processing workflow. Our findings provide a critical optimization for the operational deployment of thermal drone technology, enabling more consistent and accurate monitoring of cotton water status, which is a cornerstone for developing intelligent, site-specific irrigation management systems in arid and semi-arid agricultural regions.