Where Does Salt Live in the Landscape? Learned Context Scale for Sparse-Label Salinity Mapping

Satellite maps of soil salinity are built almost entirely from pixel-by-pixel regressors. Methods such as support vector machines, random forests, gradient-boosted trees and extreme learning machines read the spectrum at one location and predict salt content there, ignoring everything around it. Salt accumulation, however, is controlled by things a single pixel cannot see. These include low-lying depressions where water collects, distance to a drainage canal or estuary, and whether a field is irrigated or not. Under a crop canopy, the soil itself is often invisible, so the useful evidence has to come from somewhere else in the scene. The obvious fix is to add a deep network with a wide field of view. With only 100 to 500 soil samples per study area, a large network will overfit, and a skeptic will say that feeding neighborhood averages at a few radii into a tabular model would do just as well.

This proposal tests a narrower and falsifiable idea. We build a small model that looks at the landscape around each sample at several fixed distances, from about 10 m out to about 2 km. It then learns, separately for every location, how much weight to give each distance, instead of using one blend everywhere. The scientific claim is that the right context distance for salinity varies systematically across the landscape. It should be local on bare saline crusts and broad under dense vegetation or near hydrological features. A per-location choice of distance should therefore beat both pixel-only models and the same multi-distance information handed to a tabular regressor. If that fails, the negative result is itself useful: it tells soil scientists that cheap multi-radius features are sufficient.

Core Claim

Two threads of prior work motivate the proposal. The first is salinity mapping that ignores space. Soil Salinity Mapping Using Machine Learning Algorithms with the Sentinel-2 MSI in Arid Areas, China fits pixel-wise SVM, random forest and small neural networks on screened Sentinel-2 indices. Strategies for Soil Salinity Mapping Using Remote Sensing and Machine Learning in the Yellow River Delta hand-builds nine covariate strategies, including phenological "vegetation type factors," for a support vector regressor. Soil Salinity Inversion in Yellow River Delta by Regularized Extreme Learning Machine Based on ICOA tunes a single-hidden-layer network. All three treat each pixel as an isolated row of features.

Where spatial structure does enter, it enters outside the model. Monitoring Soil Salinity Classes through Remote Sensing-Based Ensemble Learning Concept: Considering Scale Effects uses Kriging only to turn point samples into 30 m pixel labels, and reports intra-region salinity variation above 119%. Soil Salinity Inversion in Coastal Corn Planting Areas by the Satellite-UAV-Ground Integration Approach bridges drone and satellite resolution with linear trend surfaces plus interpolated residuals.

The second thread is vision work showing that learned choices among several receptive-field sizes help when targets vary in scale. The Multiscale Attention Network in A Novel Attention Fully Convolutional Network Method for Synthetic Aperture Radar Image Segmentation weights parallel kernels with channel and spatial attention. The strip and dilated attention of DPSDA-Net captures elongated context in road maps. The parallel-kernel module in DRPU-YOLO11 separates small rice panicles from cluttered field backgrounds. These methods supply the mechanism, but each was trained on thousands of densely labeled images for segmentation or detection. None addresses continuous regression from a hundred point labels. DPSDA-Net's authors also note that their added modules make the network much heavier, which rules out copying those architectures directly.

Our model differs in three ways. It has almost no parameters beyond the scale-selection step. It is trained on sparse points rather than dense masks. Its scale weights are treated as a scientific output, to be checked against terrain and vegetation, rather than as an internal detail.

Research Program

For each sample location, the model receives multi-band features pooled over a small set of context sizes $r_1,\dots,r_K$, with roughly $K=5$ radii spanning 10 m to 2 km. The per-pixel features are Sentinel-2 or Landsat bands, standard salinity and vegetation indices, and elevation-derived terrain descriptors. The context at the largest radii is computed cheaply over the whole tile. We use a Fourier-domain gating layer in the style of Butterworth as Attention, whose cost grows as $N\log N$ in the number of pixels $N$. This avoids the quadratic memory cost of full non-local attention that CCNet documents. The CCNet paper's own criss-cross variant serves as an alternative global operator.

A tiny gating network reads the center pixel only. It outputs weights $\alpha_k(x)$ that sum to one, and the prediction is

$$\hat{y}(x) = g\Big(\sum_{k=1}^{K} \alpha_k(x)\, h_k(x)\Big),$$

where $h_k(x)$ is the context summary at radius $r_k$ around location $x$ and $g$ is a small regressor. The gate must decide from what it sees at the sample point how far to look. Because the gate reads only the center pixel, its behavior can be interpreted. We regularize with an entropy penalty and use strong weight decay, so the total number of trainable parameters stays in the low hundreds.

We add a second tier only if the first tier shows a gain. That tier replaces the handcrafted pixel features with an encoder pretrained without labels on unlabeled tiles from the same regions. It asks whether learned features add anything beyond indices.

Evaluation uses spatially blocked cross-validation throughout. Samples are grouped into geographic blocks larger than the largest context radius, and whole blocks are held out. Random or Kennard-Stone splits leak neighborhood information and would inflate any context-using model.

We use four field datasets described in the literature. The first is Yellow River Delta Sentinel-2 with 105 samples (Strategies…). The second is the Kongterik arid reserve with 160 samples (Soil Salinity Mapping…). The third is the Yichang/Hetao Landsat 8 set with 488 samples (Monitoring Soil Salinity Classes…). The fourth is the coastal corn set with 142 cores and drone imagery (Satellite-UAV-Ground…). Data access for these sets must be negotiated with the authors. A fallback is collecting a comparable set from Sentinel-2 plus public terrain data, using whichever field campaign becomes available.

Baselines

The naive reference is a spatial-mean model and a pixel-only regressor. For the regressor we use the best tabular model each source paper reports, refit under our blocked splits.

The published numbers show that these regressors are not weak, though they were obtained under non-spatial splits. In the Yellow River Delta, support vector regression with Strategy IX reached test $R^2 = 0.62$ and RMSE 0.38 g/kg (Strategies…). In the Kongterik reserve, an SVM on VIF-screened Sentinel-2 features reached verification $R^2 = 0.88$ and RMSE 4.89 dS m⁻¹, while a random forest reached only $R^2 = 0.27$ (Soil Salinity Mapping…). On the Yichang data, the Kriging-plus-ensemble classifier reached overall accuracy 0.85 and Kappa 0.80 (Monitoring…). On the coastal corn data, the TsHARP-plus-PLS integration reached overall $R^2 = 0.716$ (Satellite-UAV-Ground…). We expect these figures to drop under blocked splits. We report both our reproduction under their original split and under ours, so the comparison is interpretable.

The strongest competitor is a multi-radius tabular model. It receives exactly the same $h_1,\dots,h_K$ context summaries, concatenated, and fits them with XGBoost and with SVR. It sees the same information as our model and differs only in lacking a per-location choice of scale. It is also what a pragmatic soil scientist would build. A second strong competitor is regression Kriging: the pixel-wise regressor plus an ordinary Kriging model of its residuals. This is the geostatistical standard that the salinity literature already trusts.

The key ablations isolate the gate. The first fixes one global set of weights shared by all locations. The second uses a single best radius chosen by cross-validation. The third plugs a small patch CNN with a fixed receptive field into the same head.

Every model gets the same folds, the same features, and the same hyperparameter search budget of equal trial counts via nested cross-validation. Beating the multi-radius tabular model means a lower RMSE on at least three of the four datasets, with a paired block-bootstrap interval that excludes zero. Anything less counts as a tie, which we will report as such.

Why not just…?

Why not just feed multi-radius neighborhood averages into XGBoost or SVR? This is the real threat, and the proposal is built around it. Tree ensembles can in principle learn scale-dependent interactions. With 100 to 500 samples, however, the evidence shows they overfit badly: random forest's $R^2$ of 0.27 on Kongterik (Soil Salinity Mapping…), and the Yellow River Delta authors' explicit warning about random forest overfitting at small sample sizes. A single gate that shares its logic across scales has fewer ways to overfit. If the concatenated-feature baseline ties or wins under blocked cross-validation, we will conclude that learned scale selection is unnecessary and recommend the tabular recipe.

Why not just use regression Kriging? Kriging models spatial correlation of the leftover errors but assumes that correlation is the same everywhere. That assumption fails across sharp field boundaries and irrigation contrasts. The field study in Field-Scale Characterization… shows surface salinity swinging from mostly saline in the growing season to 99% non-saline after winter irrigation. Kriging also needs dense nearby samples at prediction time, while our model predicts from imagery alone. We will report where regression Kriging wins, likely in densely sampled blocks, and where it fails, likely in blocks far from training data.

Why not just train a U-Net or patch CNN? A standard encoder-decoder has one receptive field determined by its depth. It has far more parameters than 100 labels can constrain, and the vision results it would draw on, such as SegNet trailing the attention network in A Novel Attention Fully Convolutional Network…, come from dense-label regimes. The fixed-receptive-field CNN ablation tests this directly on our data.

Decisive Experiments

The first test is whether the per-location gate beats the concatenated multi-radius tabular model under spatially blocked cross-validation. If it does not, the architectural claim is dead.

The second test checks the scientific claim about the gate independently of accuracy. We correlate the weight placed on large radii with NDVI, with elevation relative to local minima, and with distance to mapped canals and coastline. Our prediction is that heavy canopy shifts weight toward broad context, because under vegetation the salt signal must come from surrounding bare ground and terrain. A gate whose weights look random would undercut the interpretation even if accuracy improves.

The third test measures transfer. We train on one region and test on another, for example the Yellow River Delta and the coastal corn area, both of which are coastal. The question is whether learned scale preferences transfer better than tabular fits.

The fourth test applies to the coastal corn set. We replace the TsHARP trend surface with the gated model applied to Sentinel-2, then check whether the 10 m map agrees better with drone-scale inversions within held-out blocks than the published $R^2 = 0.716$ pipeline does.