Conservation planners on steep farmland choose among hedgerow layouts, check dams, and vegetation restoration schemes almost entirely by how much soil each one keeps on the slope. Whether residents and visitors will accept the result is decided later, by questionnaire or not at all. One field study suggests the two criteria may conflict. Surveying check-dam landscapes, Landscape Aesthetics of Check Dams Based on Scenic Beauty Estimation Method and Artificial Intelligence Technology found that people liked vegetation and water and disliked bare land. They also disliked abundant vegetation when it looked unorganized. On the erosion side, the dense, multi-layered, irregular cover that comes back on restored land is what gives soil its cohesion and root mass. Soil detachment by overland flow under different vegetation restoration models in the Loess Plateau of China measured rill erodibility on farmland at 0.1164 s/m, against 0.0013 to 0.0039 s/m on restored lands. If people prefer tidy vegetation and messy vegetation holds soil better, there is a "tidiness penalty": the landscapes that protect soil best are rated as less beautiful.
We propose to measure this penalty directly. The plan pairs ground-level photographs of plots whose erosion has been quantified physically with scenic-beauty ratings and short written reasons from raters. The written reasons go through a text-cleaning step, modeled on a known technique for stripping stylistic padding, that turns each free-form comment into a few standardized claims such as "vegetation looks chaotic" or "bare soil visible." These claims tell us which visual cues drive the ratings, and they also serve as extra training signal for an image scorer adapted to conservation landscapes. The scientific output is the sign and size of the tidiness penalty once vegetation amount is held fixed. The methodological output is an aesthetic scorer for a domain that computational aesthetics has never covered, and it explains its scores in terms of cues that can be checked against erosion quantities.
Core Claim
Holding vegetation cover fixed, visual orderliness correlates negatively with measured soil protection. A rating model supervised with de-styled rater rationales can recover which visual cues carry that penalty more reliably than area ratios alone. We write the test as a regression over sites $i$:
$$S_i = \alpha + \beta\, \mathrm{SC}_i + \gamma\, \mathrm{FVC}_i + \delta^\top z_i + \varepsilon_i,$$
Here $S_i$ is the mean scenic-beauty score of site $i$. $\mathrm{SC}_i$ is soil conservation, defined as in Regulated Ecosystem Services Trade-Offs as potential (bare-soil) erosion minus actual erosion. $\mathrm{FVC}_i$ is fractional vegetation cover. $z_i$ collects nuisance covariates: presence of water, sky fraction, and photo conditions. $\varepsilon_i$ is noise. The tidiness penalty is a reliably negative $\beta$ that shrinks toward zero when an orderliness measure is added to $z_i$. That pattern would mean orderliness mediates the conflict. A positive or null $\beta$ refutes the hypothesis, and that result is useful too: it would tell planners that protection and appeal can be optimized together.
How This Differs From What Exists
The closest work is the check-dam study above. It regressed questionnaire scores on the area proportions of five segmented classes (vegetation, structure, sky, land, water). It computed no erosion quantity, trained no automated scorer, and recorded no reasons. Its "unorganized vegetation" finding is therefore an observation without a mechanism. That observation is our starting hypothesis, and supplying the missing erosion measurements is our first difference.
Two erosion studies supply spatial-pattern tools without any link to perception. Landscape Pattern of Sloping Garden Erosion Based on CSLE and Multi-Source Satellite Imagery computed patch fractal dimension and aggregation index over sloping gardens, with a steep-slope loss equation validated at $R^2=0.92$ against 24 field sample units. Effects of Different Hedgerow Patterns measured erodibility across four hedgerow designs and a barren control, but applied fractal dimension only to soil particle sizes. We reuse these sites and their physical measurements, and we ask whether the patterns they describe are seen and disliked.
On the aesthetics side, three methods supply our machinery. VILA showed that user comments are richer supervision than scalar scores. It trained on raw photography-forum text with no cleaning. When Attributes Disagree (AGREE) decoupled five low-level photographic attributes to avoid conflicting gradients, and explicitly left high-level composition and scene logic unmodeled. Aes-R1 trained a multimodal model to reason about aesthetics using filtered chain-of-thought data on photography benchmarks. All three stay within consumer photography and art. We move the target to conservation landscapes and replace photographic attributes with scene-level ones such as orderliness, bare-soil visibility, vegetation continuity, and how much the structures dominate the view.
The text step comes from Benchmarking LLMs' Judgments with No Gold Standard. That work compresses free-form reviews into short standardized points before scoring them with mutual information, which made its measure resistant to padding that inflated a GPT-4o judge. We use the same compression for a different purpose. Raters' reasons vary wildly in length and eloquence, and without normalization, articulate raters would dominate the attribute supervision. The known risk constrains the design: aggressive compression can erase hedging and emotional nuance, so we test for that loss explicitly rather than assume it away.
Research Program
The work has three parts. First, we build a dataset. Ground-level photographs are taken under a fixed protocol, with set viewpoint heights, times of day, and several angles per site, at locations where erosion has already been quantified. These include the Miyun hedgerow plots with measured erodibility, the six Loess Plateau land-use plots with measured detachment capacity, and a sample of the Xishuangbanna garden units with loss-equation estimates. We add check-dam scenes from the same regions. For each site we compute soil conservation and vegetation cover from the published factor maps or plot measurements. A rater panel drawn from local residents and outside visitors then gives each photo a scenic-beauty score and writes one or two sentences of reasoning. We keep the two groups separate because aesthetic judgment is known to be culturally contingent, and web-trained raters reflect narrow populations (Tikhonov et al.).
Second, we build the text pipeline. An LLM rewrites each rationale into atomic claims, and the claims are mapped onto a small fixed vocabulary of scene attributes, each with a polarity. We record claims that fall outside the vocabulary rather than discard them. Each rater's informativeness is estimated with the mutual-information score from the GEM paper, so that rationales carrying almost no information are down-weighted rather than silently kept.
Third, we build the scorer. We start from VILA's pretrained image-text model and add one adapter per attribute. Each adapter is anchored to the text of its attribute, in the style of AGREE's semantic anchors, and a ranking head is trained on the scenic-beauty scores. The attribute heads are trained on the compressed claims, and the overall head on the scores. The trained attribute outputs, especially orderliness, then become the mediator in the regression above. This lets us fit the regression on many more photographs than raters could ever score by hand.
Baselines
The naive reference points are the mean rating and the area-ratio regression from the check-dam study, refit on our photos with the same segmentation classes. Any learned scorer must clearly beat that regression, because it is exactly what current practice uses.
The strongest automated competitor is VILA's ranking adapter fine-tuned directly on our scenic-beauty scores with no rationale supervision. It reports a Spearman rank correlation of 0.774 on the AVA photography benchmark (VILA), and its zero-shot variant reports 0.657. We found no result on Lacuna for any automated model on check-dam or erosion-landscape scenic beauty, so our own zero-shot VILA run is the first reference number for the domain. VILA gets the same images, the same splits, and the same hyperparameter search budget as our model.
A pattern-metric baseline adds fractal dimension and aggregation index, computed on segmentation masks as in Landscape Pattern of Sloping Garden Erosion, to the area-ratio regression. This baseline gives an alternative, text-free measure of orderliness.
An MLLM baseline prompts a multimodal model in the Aes-R1 reasoning style to give both a score and reasons. Its reasons are also mapped onto our attribute vocabulary.
Three ablations isolate the key ingredient. We compare training on raw rationales, on compressed claims, and on no text at all, while holding the architecture fixed.
"Beating" a baseline has two parts. On prediction, it means a higher Spearman correlation on held-out sites, as opposed to held-out photos of the same site, with bootstrap confidence intervals that exclude the baseline. On explanation, it means the attributes the model assigns match raters' compressed claims on held-out photos better than the baseline's attributes do.
Why not just…?
Why not just fine-tune VILA on the scenic-beauty scores? At a few hundred sites this may well reach competitive rank correlation. It produces a single number, though, so it cannot separate orderliness from greenness, and the science question depends on that separation. The comparison in the Baselines section shows whether rationale supervision costs anything in accuracy. If it does not, the attribute heads come for free.
Why not just add fractal dimension and aggregation to the area-ratio regression? This is the real threat. It may capture "unorganized vegetation" without any text, and if it predicts ratings as well as our scorer, the prediction contribution of the text pipeline disappears. We therefore make the penalty estimate the primary result and treat the scorer as one of two orderliness instruments. If pattern metrics and the learned orderliness head give the same $\beta$ attenuation, that agreement confirms the finding. The rationales still tell us whether raters describe what the metrics measure. Mask geometry may miss visual disorder such as dead stems or mixed species.
Why not just ask a multimodal LLM to rate and explain? Aesthetic rationales generated by these models are more error-prone than factual descriptions (Xiao et al.), and existing reasoning models are trained on photography. We run this baseline and measure how well its reasons agree with human claims. We expect agreement to be weakest on the erosion-specific cue of visible bare soil.
Why not just score raw rationales with an LLM judge instead of compressing them? The GEM paper showed that LLM judges inflate scores for padded text, while the compressed pipeline stayed stable. Its compressed score reached 0.479 Spearman against instructor grades, versus 0.537 for the GPT-4o judge. That comparison shows compression gives robustness at some cost in raw correlation, and our raw-versus-compressed ablation measures the same trade-off in this domain.
Decisive Checks
The central check is the sign of $\beta$ within strata of vegetation cover, and whether it attenuates when orderliness is added. We run it separately for residents and visitors, and separately for the Loess restoration plots, where the physical contrast is largest. A second check is a within-site counterfactual. For the Miyun plots, the lowest-erodibility hedgerow design should score as less orderly than the others if the penalty is real. A third check tests information loss: we sample compressed claims, have annotators compare them with the original rationales, and record how often hedged or emotional content is dropped. If more than a small fraction is lost, we stop relying on the compressed text for the attribute heads. A negative $\beta$ that survives all three checks would give planners the first quantitative estimate of how much visual acceptance effective erosion control costs, and the attribute outputs would indicate which layout changes could win back acceptance without giving up protection.
References
- Landscape Aesthetics of Check Dams Based on Scenic Beauty Estimation Method and Artificial Intelligence Technology
- Soil detachment by overland flow under different vegetation restoration models in the Loess Plateau of China
- Regulated Ecosystem Services Trade-Offs: Synergy Research and Driver Identification in the Vegetation Restoration Area of the Middle Stream of the Yellow River
- Landscape Pattern of Sloping Garden Erosion Based on CSLE and Multi-Source Satellite Imagery in Tropical Xishuangbanna, Southwest China
- Effects of Different Hedgerow Patterns on the Soil Physicochemical Properties, Erodibility, and Fractal Characteristics of Slope Farmland in the Miyun Reservoir Area
- VILA: Learning Image Aesthetics from User Comments with Vision-Language Pretraining
- When Attributes Disagree: Gradient Conflict in Image Aesthetic Assessment
- Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization
- Benchmarking LLMs' Judgments with No Gold Standard
- Exploring Cognitive and Aesthetic Causality for Multimodal Aspect-Based Sentiment Analysis