Monitoring Spatial Contrasts in Soil Total Salt Content Among Deep-Tillage Management Zones Using Low-Cost RGB-NIR Sensing and Optimized Machine Learning

Soil salinization restricts crop production and complicates field management in arid farming systems. We assessed low-cost red-green-blue and near-infrared sensing across three contiguous management zones assigned to deep vertical rotary tillage depths of 40, 55, and 70 cm in Bachu County, Xinjiang, China. The dataset contained 90 paired sampling points, with 30 spatial subsamples in each zone. Because each depth was represented by one zone, the design had no independent plot-level biological replication and zone contrasts were interpreted descriptively. We compared support vector regression and extreme gradient boosting after feature screening by the maximal information coefficient, Pearson correlation, or grey relational analysis. Hyperparameters were tuned by conventional random search, the Sparrow Search Algorithm, or the Grey Wolf Optimizer within repeated nested cross-validation. Mean soil total salt contents were 12.12, 14.63, and 7.24 g kg−1 in the 40, 55, and 70 cm zones, respectively. Welch’s analysis of variance and Games–Howell comparisons indicated that the 70 cm zone differed from the other two zones, whereas the 40 and 55 cm zones did not differ. These point-level tests describe spatial contrasts and do not establish a causal tillage-depth effect. Using field-derived RGB-NIR predictors, the best screened and conventionally tuned extreme-gradient-boosting model achieved a pooled out-of-fold coefficient of determination of 0.694 and a root mean square error of 3.832 g kg−1. Its paired outer-fold error was lower than that of the all-feature default baseline after multiplicity correction, while the two swarm optimizers did not significantly improve on conventional tuning under the matched evaluation budget. The results support field-derived RGB-NIR sensing as a practical aid for within-field monitoring of soil total salt content, while non-spatial cross-validation may overestimate performance at spatially independent locations and external validation remains necessary.