ECHO is a context-calibrated regression framework for estimating body-weight-normalized feed intake of Pacific white shrimp (Litopenaeus vannamei) from passive acoustic observations. It combines a log-Mel acoustic sequence with four context variables: nominal shrimp body-weight level, ammonia nitrogen concentration, nitrite nitrogen concentration, and stocking density.
Context is introduced through:
- state-level calibration before selective state-space sequence modeling; and
- feature-level gating after attention-based temporal aggregation.
The implementation is intended for:
- research on passive acoustic monitoring in aquaculture;
- replication and extension of context-conditioned regression methods;
- ablation and baseline comparisons using appropriately collected data.
The repository is not intended to:
- autonomously control commercial feeding equipment without independent validation and safeguards;
- diagnose animal health or water quality;
- generalize to species, sensors, facilities, or environmental ranges that were not represented during training;
- replace direct measurements in regulatory or welfare-critical decisions.
The neural models consume:
- a fixed-length
[64, time]log-Mel tensor; and - a standardized four-dimensional context vector.
The output is a scalar estimate of body-weight-normalized feed intake in g/g.
The original feeding-trial recordings and labels are not distributed in this repository. Users are responsible for ensuring that replacement data can be lawfully collected, stored, and released.
The manuscript specifies the principal preprocessing, calibration equations, latent dimension, sequence depth, and training protocol. Low-level channel widths and estimation-head widths are exposed as reference defaults because they are not uniquely fixed by the manuscript text. Exact numerical reproduction requires the original data, archived condition split, and exact experiment configuration.
- A 200-s window inherits a trial-level target measured over the complete feeding observation, so temporal label precision is limited.
- The target distribution is right-skewed and contains a sparse high-intake tail.
- Context variables can be correlated with experimental treatment design.
- Condition-level splitting reduces leakage but does not establish generalization to independent farms or sensors.
- The pure-PyTorch selective-SSM fallback is slower and is not numerically
identical to the fused
mamba-ssmimplementation.
Use condition-level splits and report at least R2, MAE, and RMSE. Claims about model differences should use paired predictions and a resampling unit that preserves the relevant experimental grouping.