EmissionNet-Transformer: Forecasting Agricultural Emissions

Updated 6 July 2026

The paper presents a hybrid convolutional-transformer that leverages convolution for local feature extraction and transformer attention for long-range dependency modeling in monthly N2O emission forecasts.
ENT uses a 24-month context window of multi-channel global emissions maps from EDGAR to capture seasonal patterns and spatial variations effectively.
The model demonstrates parameter efficiency and stability in autoregressive rollout, achieving lower test MSE than ConvLSTM and MLP, though ENV outperforms in one-step accuracy.

Searching arXiv for the primary paper and closely related acronym collisions to ground the article in current arXiv records. arXiv search query: (Saligram et al., 7 Jul 2025) EmissionNet: Air Quality Pollution Forecasting for Agriculture EmissionNet-Transformer (ENT) is the transformer-augmented model introduced in “EmissionNet: Air Quality Pollution Forecasting for Agriculture” for forecasting next-month agricultural $\mathrm{N_2O}$ emissions from historical global emissions maps. In the reported formulation, the model consumes a 24-month context of monthly emissions for five molecules and predicts the spatially resolved $\mathrm{N_2O}$ map for month $t+24$ , making ENT a hybrid convolutional-transformer spatiotemporal forecaster tailored to agricultural emissions and air-quality analysis rather than a generic vision architecture (Saligram et al., 7 Jul 2025).

1. Forecasting problem and scientific rationale

ENT is defined around a one-step monthly forecasting task on global gridded agricultural emissions. The paper formalizes the input sequence as

$X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$

with prediction target

$x_{t+24},$

where the target is the next month’s $\mathrm{N_2O}$ emissions map (Saligram et al., 7 Jul 2025).

The motivation is domain-specific. The study frames agricultural air-quality forecasting as a setting in which standard physics-based models such as CMAQ, AERMOD, and NASA GEOS-CF remain valuable but can struggle with complex nonlinear pollutant interactions. Within machine learning, the comparison is made against an MLP baseline, which lacks strong spatial and temporal inductive bias, and ConvLSTM, which captures spatiotemporal structure but is limited by recurrent modeling and local receptive fields. ENT is proposed because agricultural emissions are described as having strong seasonal periodicity, high autocorrelation, spatially structured global patterns, and a simultaneous need to represent local variation and long-range dependencies (Saligram et al., 7 Jul 2025).

The model’s intended role follows directly from that framing. Convolutional layers are used to extract local spatial features, while transformer attention is used to model long-range dependencies across the spatial grid. The paper explicitly connects this design to a qualitative failure mode observed in ConvLSTM: high-gradient continental boundaries are captured more effectively than oceans or other low-gradient regions, where ConvLSTM exhibits an almost constant offset. ENT is positioned as a mechanism for mitigating that localized bias through global attention (Saligram et al., 7 Jul 2025).

2. Data source, preprocessing, and benchmark construction

The benchmark is derived from the EDGAR global greenhouse-gas emissions database. Raw data are annual netCDF files with 12 monthly slices per year spanning 2000–2023, with global latitude coverage $[-90^\circ,+90^\circ]$ , longitude coverage $[-180^\circ,+180^\circ]$ , and native $0.1^\circ$ spatial resolution. The raw annual files are reported to occupy about 72 GB and contain emissions and flux maps for five molecules: $\mathrm{CH_4}$ , $\mathrm{N_2O}$ 0, $\mathrm{N_2O}$ 1, $\mathrm{N_2O}$ 2, and GWA (Saligram et al., 7 Jul 2025).

A direct model over the original concatenated tensor is described as intractable for the study’s purposes; the authors estimate roughly 480M parameters for a direct multilayer neural network. Two preprocessing operations are therefore applied. First, the spatial grid is pooled from $\mathrm{N_2O}$ 3 to $\mathrm{N_2O}$ 4. Second, the flux dimension is discarded because flux prediction is not the task. After preprocessing, the final tensor shape is

$\mathrm{N_2O}$ 5

with 5 molecular emission channels, 288 monthly steps, and a $\mathrm{N_2O}$ 6 pooled latitude-longitude grid. Each monthly observation is thus a 5-channel global map, and the predictive target is a single-channel $\mathrm{N_2O}$ 7 map one month ahead (Saligram et al., 7 Jul 2025).

The data split is chronological, and the paper uses a rolling-window setup with a 24-month context window and 1-month forecast horizon.

Split	Period	Months
Training	January 2000 to March 2019	231
Validation	April 2019 to July 2021	28
Test	August 2021 to January 2024	29

This design makes the benchmark explicitly out-of-time rather than randomly shuffled, which is appropriate for emissions forecasting. The paper does not report MAE, $\mathrm{N_2O}$ 8, correlation, uncertainty intervals, or region-specific skill metrics; its primary quantitative metric is mean squared error, with additional discussion of RMSE growth under autoregressive rollout (Saligram et al., 7 Jul 2025).

3. Architectural composition of ENT

ENT is described as a hybrid convolutional-transformer spatiotemporal forecasting model composed of an initial convolution, three residual blocks, a lightweight FlashTransformerDecoder attention module, a final $\mathrm{N_2O}$ 9 convolution, and bilinear upsampling (Saligram et al., 7 Jul 2025).

A concise summary of the paper-stated components is given below.

Component	Paper-stated specification	Role
Initial stem	$t+24$ 0 convolution, 64 filters, batch normalization, ReLU	Low-level spatial feature extraction
Residual trunk	Three residual blocks	Downsampling, deeper feature extraction, reconstruction
Attention module	Lightweight FlashTransformerDecoder with self-attention, linear $t+24$ 1, dropout, layer normalization	Long-range dependency modeling
Output head	$t+24$ 2 convolution + bilinear upsampling	Final spatial prediction

The initial operation is a $t+24$ 3 convolution with 64 filters, followed by batch normalization and ReLU. Residual block 1 applies a $t+24$ 4 convolution with stride 2 to downsample, followed by a second $t+24$ 5 convolution, with a $t+24$ 6 convolutional downsampling layer in the shortcut path to match dimensions. Residual block 2 maintains spatial resolution while continuing deeper feature extraction. Residual block 3 is described as reconstructing the feature map and reducing the number of channels back to 64. Intermediate channel counts and exact activation ordering inside the residual stack are not reported (Saligram et al., 7 Jul 2025).

After convolutional encoding, ENT applies a lightweight FlashTransformerDecoder module. The paper states that this module performs self-attention over spatial embeddings, uses linear projections for queries, keys, and values, and includes dropout and layer normalization. Its stated purpose is to capture long-range dependencies between emissions. However, several implementation-critical details are absent: the number of heads, hidden dimension, feed-forward dimension, decoder depth, tokenization or flattening scheme, whether cross-attention is used, and whether “Flash” refers to a specific FlashAttention implementation are not specified (Saligram et al., 7 Jul 2025).

The paper is similarly incomplete regarding spatiotemporal encoding. It does not provide a positional encoding formula, a learned positional embedding description, or any explicit temporal encoding mechanism for ENT. It is therefore only possible to state that the model consumes a 24-month context window with five emission channels per month; the internal arrangement of the 24 time steps in the ENT input tensor is not specified (Saligram et al., 7 Jul 2025).

After transformer refinement, features are projected through a $t+24$ 7 convolution and then passed to a bilinear upsampling layer said to restore spatial resolution to $t+24$ 8. This creates a documented inconsistency with the dataset section, which reports a preprocessed grid of $t+24$ 9. The paper does not resolve whether ENT operates and is evaluated at $X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 0, whether additional downsampling is performed elsewhere, or how that resolution relates to the benchmark tensor (Saligram et al., 7 Jul 2025).

4. Optimization protocol and evaluation methodology

The selected ENT configuration in the hyperparameter sweep is ENT_3e-3_16_24_1_200_AdamW, with learning rate $X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 1, batch size 16, context window 24, forecast horizon 1, 200 epochs, optimizer AdamW, and test MSE $X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 2. A second reported ENT variant, ENT_5e-4_32_24_1_500_Adam, reaches test MSE $X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 3 (Saligram et al., 7 Jul 2025).

The broader training regime is described as using hyperparameter sweeps that identified dynamic learning rates of $X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 4, warmup ratios of 0.2–0.3, and weight decay of $X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 5, with training run for 200 epochs after which improvements were negligible. Batch sizes were reduced from 32 to 8 for more complex models. For ENT specifically, the paper explicitly mentions batch normalization after the initial convolution, along with dropout and layer normalization in the transformer module. The exact warmup ratio for ENT, dropout rate, and whether early stopping was used are not reported (Saligram et al., 7 Jul 2025).

Evaluation is centered on MSE. The reported results tables contain train, validation, and test MSE; the paper also includes a figure titled “Auto-Regressive RMSE over Time,” but does not provide an explicit RMSE formula. Likewise, no explicit training-loss equation is written for ENT, even though the results section indicates that MSE is the primary optimization target. Hardware details are largely absent: the authors acknowledge access to compute for training and evaluation but do not specify GPU model, CPU, memory, training time, or framework-level system configuration (Saligram et al., 7 Jul 2025).

One further methodological ambiguity concerns optimizer reporting. The main text states that all models used “the same adaptive learning rate algorithm,” while the appendix sweep table lists different optimizers across models, including AdamW, Adam, and AdaGL. The paper does not reconcile that discrepancy (Saligram et al., 7 Jul 2025).

5. Empirical performance and comparative position

On one-step next-month forecasting, ENT is the second-best model in the study by test MSE, clearly outperforming MLP and ConvLSTM while remaining behind the heavier convolutional ENV model (Saligram et al., 7 Jul 2025).

Model	Test MSE	Parameters
MLP	0.06288	$X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 6M
ConvLSTM	0.00156	$X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 7M
ENT	0.00068	$X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 8M
ENV	0.00010	$X=\{x_t, x_{t+1}, \dots, x_{t+23}\},$ 9M

These numbers establish three central comparative facts. First, ENT reduces test MSE from $x_{t+24},$ 0 to $x_{t+24},$ 1 relative to ConvLSTM. Second, it improves over MLP by a much larger margin, from $x_{t+24},$ 2 to $x_{t+24},$ 3. Third, it does not surpass ENV’s one-step test MSE of $x_{t+24},$ 4. The study therefore distinguishes between absolute next-month accuracy, where ENV leads, and the particular strengths of ENT, namely parameter efficiency and autoregressive stability (Saligram et al., 7 Jul 2025).

Parameter count is especially important in the paper’s interpretation. ENT is the smallest model in the benchmark at approximately $x_{t+24},$ 5M parameters, yet it outperforms ConvLSTM and MLP and approaches ENV much more closely than its size would suggest. This suggests good computational efficiency relative to the baselines and to ENV, although explicit runtime measurements are not provided (Saligram et al., 7 Jul 2025).

Learning dynamics also differentiate ENT from the simpler baselines. The paper states that ENT and ENV continue learning much longer than MLP and ConvLSTM, with ENT/ENV beginning to converge after about 150 epochs, whereas MLP/ConvLSTM converge around 10 epochs. On a logarithmic MSE scale, ENT reaches substantially lower error than the simpler models (Saligram et al., 7 Jul 2025).

The qualitative results align with the quantitative picture. In map comparisons across forecast offsets $x_{t+24},$ 6, where each unit of $x_{t+24},$ 7 is one month, ENT is described as improving on ConvLSTM and MLP while avoiding the oceanic or low-gradient offset bias visually associated with ConvLSTM. The model is said to capture both high-contrast continental regions and smoothly varying open-water areas more faithfully (Saligram et al., 7 Jul 2025).

The paper’s most favorable discussion of ENT concerns long-horizon autoregressive rollout. All models accumulate error when predictions are recursively fed back as inputs, but ENT is reported to approach an asymptote, whereas MLP, ConvLSTM, and ENV continue growing. The authors attribute this to transformer attention capturing long-range dependencies that help stabilize repeated forecasting. Under the paper’s own interpretation, ENV is preferable when the objective is the best one-step next-month MSE, while ENT is preferable when parameter efficiency and stability under autoregressive multi-step rollout are more salient criteria (Saligram et al., 7 Jul 2025).

6. Limitations, deployment implications, and nomenclature

ENT’s methodological description is intentionally high-level, and that leaves several open technical questions. The paper does not provide a layer-by-layer tensor-shape specification, exact temporal tokenization scheme, explicit self-attention or multi-head-attention equations for ENT, positional or spatiotemporal encodings, decoder depth, attention-head count, or feed-forward dimensionalities. It also omits uncertainty quantification, region-level attribution, attention visualization, saliency analysis, and operational deployment details such as latency budgets or data-assimilation interfaces (Saligram et al., 7 Jul 2025).

Several textual inconsistencies are also documented. The most visible is the mismatch between the preprocessed benchmark grid of $x_{t+24},$ 8 and the statement that bilinear upsampling restores output resolution to $x_{t+24},$ 9. A second is the optimizer inconsistency between the main text and appendix. A third appears in the conclusion, where transformer and flash-attention benefits are attributed to ENV even though those architectural elements belong to ENT in the methods section. A faithful reading therefore relies primarily on the methods and results descriptions when characterizing ENT (Saligram et al., 7 Jul 2025).

Despite those limitations, the deployment relevance claimed for ENT is clear. The model is built for monthly forecasts of spatially resolved agricultural $\mathrm{N_2O}$ 0 emissions on global grids, with the stated practical implications of monitoring seasonal emission cycles, supporting environmental policy, improving planning for agricultural management, and anticipating pollution patterns that standard physics-based systems may model less flexibly. Because the paper does not report uncertainty estimates, policy thresholds, or an operational stack, those implications remain conceptual rather than fully operationalized (Saligram et al., 7 Jul 2025).

A recurrent source of confusion is nomenclature. “ENT” in this context refers specifically to EmissionNet-Transformer from the agricultural emissions forecasting paper, and it should not be conflated with other arXiv models or methods that use similar abbreviations, including EnTransformer for multivariate probabilistic forecasting (Pathak et al., 12 Mar 2026), Emotion Neural Transducer for speech emotion recognition (Shen et al., 2024), or Error Norm Truncation for robust text-generation training (Li et al., 2023). Within the emissions-forecasting literature summarized here, ENT denotes the lightweight convolutional-transformer model for next-month agricultural $\mathrm{N_2O}$ 1 prediction from 24 months of multi-emission history.