Guanyuan Shuai

Estimating soil organic carbon from satellite-only datasets typically requires bare soil pixels, often contaminated by cloud and residue cover from conservation practices. Geospatial foundation models produce spatially continuous satellite embeddings that retain soil information, and this study evaluated their potential for SOC estimation across space and time. Using 2,003 topsoil samples from Indiana, Minnesota, and Missouri between 2021 and 2023, we compared three geospatial foundation models, AlphaEarth Foundations (AEF), Prithvi, and Tessera, against Harmonized Landsat and Sentinel-2 (HLS) spectral features. We tested sixteen modeling approaches spanning linear/distance-based, tree-based, neural network, and tabular foundation models, and assessed spatial and temporal transferability by training and testing across different states and years. AEF and Tessera outperformed HLS-derived features, and combining them with tabular foundation models such as Tabular Prior Fitted Network (TabPFN) achieved R² values of up to 0.73. Cross-year and cross-state performance was limited for all models (R² of 0.04-0.15), but incorporating a small proportion of target-year or target-state samples, through point-level sampling, improved model performance (R² up to 0.62 for cross-year and 0.54 for cross-state). Bulk density was the most important predictor identified through Shapley Additive Explanations (SHAP) analysis, and publicly available bulk density products (SSURGO and POLARIS) offered a viable, though slightly less accurate (5-10% decrease in R²), substitute for field measurements. These findings suggest that combining geospatial and tabular foundation models offers a scalable, accurate approach to SOC estimation, while highlighting spatial and temporal generalization as a continuing challenge for large-scale monitoring.

Despite the new equipment capabilities, uneven crop stands are still common occurrences in crop fields, mainly due to spatial heterogeneity in soil conditions, seedling mortality due to herbivore predation and disease, or human error. Non-uniform plant stands may reduce grain yield in crops like maize. Thus, detecting signs of variability in crop stand density early in the season provides critical information for management decisions and crop yield forecasts. Processing techniques applied on images captured by unmanned aerial vehicles (UAVs) has been used successfully to identify crop rows and estimate stand density and, most recently, to estimate plant-to-plant interval distance. Here, we further test and apply an image processing algorithm on UAV images collected from yield-stability zones in a commercial crop field. Our objective was to implement the algorithm to compare variation of plant-spacing intervals to test whether yield differences within these zones are related to differences in crop stand characteristics. Our analysis indicates that the algorithm can be reliably used to estimate plant counts (precision >95% and recall >97%) and plant distance interval (R² ~0.9 and relative error <10%). Analysis of the collected data indicated that plant spacing variability differences were small among plots with large yield differences, suggesting that it was not a major cause of yield variability across zones with distinct yield history. This analysis provides an example of how plant-detection algorithms can be applied to improve the understanding of patterns of spatial and temporal yield variability.

Unmanned aerial vehicle technology which is capable of acquiring the centimeter-level orthographic images provides a promising way to map the paddy rice extent. However, crop spectral mixture is still the major that limits the performance of paddy rice identification when using the spectral information alone. In order to eliminate the influence of spectral mixture on crop mapping accuracy, an innovative method was proposed to produce field-level canopy height data by calculating the elevation difference between the vegetated and nonvegetated parcels. We applied support vector machine on four types of datasets: 1) Pixel-based Spectral Features (PSF); 2) PSF and Canopy Height Features; 3) Object-based Spectral Features (OSF); and 4) Object-based Spectral Features and Canopy Height Features (OSCHF). The results showed that the full-class classification accuracy from OSCHF was the highest with an overall accuracy of 94.04% and Kappa of 0.91, which was significantly higher than result using OSF, and the accuracy of PSF was the worst. The OSF could eliminate speckle noise problem to some extent, but grapes and trees were still to some extent misclassified as paddy rice due to similar spectrum. Fortunately, these confusions were effectively avoided by including the canopy height.

This paper developed an approach to determine optimal parameters, C and s, for support vector domain description (SVDD) model to map specific land cover from integrating of training and window-based validation sets (WVS-SVDD). The validation set based on window-based approach made a tighten hypersphere because of compact constraint by the outlier pixels which were located closely to the target class in the feature space. The target land coves of wheat and bare land were considered to test the proposed method's performance. The overall accuracy for wheat reached as high as 94.37%. However, the underestimation of wheat, only 71.12% of the user's accuracy, attributed to the validation set covering a small portion of wheat spectra. The larger window sizes were tested to achieve more wheat pixels' samples for validation set. The results showed that wheat accuracy were improved along with window size increasing and the overall accuracies were higher than 88% The bare land as more heterogeneous land cover against wheat was selected to analyze the applicability and suitability of WVS-SVDD, SVDD classification was conducted and compared to the support vector machine (SVM) method as benchmark. The producer's and user's accuracies for bare land were over 80% at 2.4-m resolution scale and the overall accuracies were similar to those produced by the SVM at coarser spatial resolutions, highlighting the applicability of WVS-SVDD. Therefore, the developed method showed its advantages using the optimal parameters, C and s, for mapping homogeneous wheat and heterogeneous bare land, which exhibits great potentials to achieve highly accurate specific land cover.

Remote sensing provides a way to identify paddy and measure the planting area with high efficiency. Particularly, the multi-temporal images show dominant advantages of extracting short-term spectral features of paddy. Traditional change detection methods usually try to derive exhaustive training dataset for classification which is not suitable for those applications that are interested on one specific class. This work introduces a method for paddy mapping by Support vector data description (SVDD) with multi-temporal HJ-1 images, at a typical study area in the middle region of Jiangsu Province. This method aims at mapping the image data into a high dimensional feature space where a hypersphere that encloses most of the patterns belonging to the class of interest (target class) and rejecting the rest (outliers) can be defined. The results showed that the proposed method could precisely identify the characteristic spectral change of paddy and efficiently map paddy with, the overall accuracy at 93.75%, which is slightly lower than that with SVM approach 95.25%. This method decreases the difficulties in training data and could be adapted for further applications.