Soil salinization critically threatens global agriculture, with approximately 20% of cultivated land affected worldwide. Traditional soil salinity monitoring methods rely on labor-intensive field sampling, providing limited spatial coverage. Remote sensing offers a promising alternative, yet satellite data alone suffers from coarse resolution (10m for Sentinel-2A), while drone technology provides centimeter-scale precision but lacks large-area coverage. We address this gap by integrating Unmanned Aerial Vehicle (UAV) and satellite data through scale transformation to achieve high-precision, large-scale soil salinity mapping.
Our study area encompassed four agricultural regions (15–20 ha each) within China’s Hetao Irrigation District, characterized by severe secondary salinization due to intensive irrigation and arid conditions. Field campaigns during April 8–12, 2023, collected 224 topsoil (0–20 cm) samples. Soil Salt Content (SSC) was derived from electrical conductivity measurements:
$$ SSC = (0.2882 \times EC_{1:5} + 0.0183) \times 100\% $$
where \( EC_{1:5} \) represents electrical conductivity at 1:5 soil:water ratio. After outlier removal, 220 samples were partitioned into training (n=150) and validation (n=70) sets (Table 1).
| Dataset | Samples | Salinity Level Distribution | SSC Statistics | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Non-saline | Slight | Moderate | Severe | Saline | Mean (%) | SD (%) | CV (%) | ||
| Overall | 220 | 68 | 38 | 51 | 40 | 23 | 0.48 | 0.39 | 81 |
| Training | 150 | 46 | 24 | 39 | 27 | 14 | 0.48 | 0.37 | 77 |
| Validation | 70 | 22 | 14 | 12 | 13 | 9 | 0.50 | 0.42 | 84 |
Simultaneously, we deployed a RedEdge-M multispectral sensor (475, 560, 668, 717, 840 nm) on a DJI Matrice 600 drone technology platform. Flights at 100m altitude yielded 7 cm resolution imagery, radiometrically calibrated using ground reference panels. Sentinel-2A imagery (April 11, 2023) was processed to surface reflectance at 10m resolution. Fifteen spectral indices were calculated for both platforms (Table 2).
| Index | Formula | Reference |
|---|---|---|
| SI1 | \(\sqrt{\text{Blue} \times \text{Red}}\) | Zhang et al. (2020) |
| Int2 | \((\text{Red} + \text{Green})/2\) | Triki et al. (2015) |
| BI | \(\sqrt{(\text{Red}^2 + \text{Green}^2)/2}\) | Khan et al. (2005) |
| SI2 | \(\sqrt{\text{Green} \times \text{Red}}\) | Allbed et al. (2014) |
| SI8 | \(\text{Blue} \times \text{Red}/\text{Green}\) | Bannari et al. (2008) |
Optimal spectral indices were selected through Pearson Correlation Coefficients (PCC) and Recursive Feature Elimination (RFE). Four machine learning models—Partial Least Squares Regression (PLSR), Random Forest (RF), Backpropagation Neural Network (BPNN), and Support Vector Regression (SVR)—were implemented for SSC inversion. Drone-derived salinity maps were resampled to 1m, 5m, and 10m resolutions using three methods:
- Nearest Neighbor (Nearest)
- Bilinear Interpolation (Bilinear)
- Cubic Convolution (Cubic)
Scale-transformed SSC values were extracted from Sentinel-2A corresponding pixels to establish satellite-scale inversion models.
PCC analysis identified SI1 as the most sensitive UAV spectral index (\(r = 0.75\)). RFE further selected five optimal features: SI1, Int2, SI2, BI, and SI8. Among UAV inversion models (Table 3), RF demonstrated superior performance (\(R^2_{train} = 0.91\), \(RMSE_{train} = 0.16\%\); \(R^2_{val} = 0.82\), \(RMSE_{val} = 0.23\%\)).
| Model | Training Set | Validation Set | ||
|---|---|---|---|---|
| R² | RMSE (%) | R² | RMSE (%) | |
| PLSR | 0.62 | 0.32 | 0.70 | 0.27 |
| RF | 0.91 | 0.16 | 0.82 | 0.23 |
| BPNN | 0.79 | 0.37 | 0.73 | 0.22 |
| SVR | 0.80 | 0.26 | 0.77 | 0.21 |
Bilinear resampling outperformed other methods across scales, showing highest correlations with Sentinel-2A spectral indices (Table 4). Satellite-scale model accuracy decreased with coarser resolutions but consistently surpassed direct field-Sentinel modeling (OSSC). The 0.07m UAV scale achieved optimal satellite inversion when coupled with SVR (\(R^2_{val} = 0.72\), \(RMSE_{val} = 0.14\%\)), improving OSSC performance by \(ΔR^2 = +0.30\) and \(ΔRMSE = -0.19\%\).
| Scale | Optimal Model | Training Set | Validation Set | ||
|---|---|---|---|---|---|
| R² | RMSE (%) | R² | RMSE (%) | ||
| OSSC | RF | 0.49 | 0.14 | 0.42 | 0.33 |
| 0.07m | SVR | 0.69 | 0.15 | 0.72 | 0.14 |
| 1m | RF | 0.69 | 0.07 | 0.65 | 0.16 |
| 5m | SVR | 0.63 | 0.23 | 0.69 | 0.22 |
| 10m | RF | 0.57 | 0.16 | 0.55 | 0.27 |
Regional salinity mapping revealed that 76.5% of Hetao croplands exhibited slight-to-moderate salinization (0.2–0.6% SSC), with severe salinization concentrated in northwestern areas and Wuliangsuhai Lake periphery. Classification accuracy assessment showed UAV-satellite synergy reduced misclassification of non-saline soils as slightly saline by 9–16% compared to OSSC-based mapping.
This research demonstrates that drone technology bridges the spatial-scale gap between point samples and satellite pixels. UAV-derived high-resolution salinity maps provide optimal training data for satellite models when resampled via bilinear interpolation. The Unmanned Aerial Vehicle platform enables scale-aware salinity representation, with 0.07m resolution achieving superior satellite inversion. Future work should integrate soil moisture and texture parameters to enhance model robustness and explore deep learning for multi-scale feature fusion. Our framework establishes a practical pathway for operational large-scale salinization monitoring using cost-effective drone-satellite synergy.