[Paper Review] A Scalable Machine Learning System for Pre-Season Agriculture Yield Forecast

I.V.P. De Oliveira, Renato L. F. Cunha+2 more|arXiv (Cornell University)|Jun 25, 2018

Remote Sensing in Agriculture18 references19 citations

TL;DR

This paper presents a scalable machine learning system for pre-season soybean and maize yield forecasting in Brazil and the USA, using satellite-derived precipitation, temperature, soil properties, and seasonal climate forecasts—without relying on NDVI data. The system achieves comparable accuracy to NDVI-dependent models, enabling early-season decision-making for farmers through a hybrid deep learning architecture combining LSTM and fully connected layers trained on historical yield data.

ABSTRACT

Yield forecast is essential to agriculture stakeholders and can be obtained with the use of machine learning models and data coming from multiple sources. Most solutions for yield forecast rely on NDVI (Normalized Difference Vegetation Index) data, which is time-consuming to be acquired and processed. To bring scalability for yield forecast, in the present paper we describe a system that incorporates satellite-derived precipitation and soil properties datasets, seasonal climate forecasting data from physical models and other sources to produce a pre-season prediction of soybean/maize yield---with no need of NDVI data. This system provides significantly useful results by the exempting the need for high-resolution remote-sensing data and allowing farmers to prepare for adverse climate influence on the crop cycle. In our studies, we forecast the soybean and maize yields for Brazil and USA, which corresponded to 44% of the world's grain production in 2016. Results show the error metrics for soybean and maize yield forecasts are comparable to similar systems that only provide yield forecast information in the first weeks to months of the crop cycle.

Motivation & Objective

• To develop a scalable, pre-season yield forecasting system that reduces dependency on high-resolution remote sensing data like NDVI.
• To enable early agricultural decision-making by providing yield forecasts before planting begins.
• To demonstrate that climate and soil data alone can support accurate yield predictions comparable to NDVI-based models.
• To create a globally scalable system using standardized, publicly available datasets with 250m spatial resolution.

Proposed method

• The system uses a hybrid deep neural network with long short-term memory (LSTM) layers for dynamic meteorological data (precipitation, temperature) and fully connected layers for static soil properties.
• Input features include historical precipitation, temperature (min/max), and soil characteristics (e.g., texture, organic content) at 250m resolution.
• The model is trained on historical yield data at the municipality level for 1,500+ cities in Brazil and the USA.
• Weather forecasts are sourced from a reanalysis-based seasonal forecast product, enabling predictions up to seven months in advance.
• The system leverages RESTful APIs to dynamically retrieve location-specific weather and soil data on demand.
• Model performance is evaluated using standard metrics: MAE, MAPE, RMSE, RMSPE, and R² across multiple regions and crops.

Experimental results

Research questions

• RQ1 Can yield forecasts be accurately generated before planting using only climate and soil data, without NDVI or in-season remote sensing?
• RQ2 How does the performance of a pre-season forecast system based on meteorological and soil data compare to NDVI-based models that operate later in the crop cycle?
• RQ3 To what extent can a deep learning model implicitly learn crop cycle patterns from seasonal climate and soil inputs alone?
• RQ4 Can such a system be scaled globally with minimal data requirements and maintain high predictive accuracy?

Key findings

• The model achieved a mean absolute percentage error (MAPE) of 10.70% for soybean yield forecasting in Brazil and 9.80% in the USA, with R² values of 0.55 and 0.75 respectively.
• For maize, the system recorded a MAPE of 11.31% in the USA and an R² of 0.71, demonstrating strong performance across crops.
• The system's performance is comparable to existing NDVI-based models, despite not requiring in-season remote sensing data.
• The hybrid architecture effectively captures temporal dependencies in climate data and spatial patterns in soil data, enabling accurate pre-season predictions.
• The model's ability to detect redundant information in input data suggests robust generalization across diverse agro-climatic regions.
• The system enables actionable pre-season decisions, such as crop type or genetic variety selection, due to its early forecast lead time.