Constructing Vegetation-Health-Map Image Forecasts Using a Novel Variable Length Attention ConvLSTM Network
CSEF · 2019 Computational Systems & Analysis (Senior Division Only)
Overview
Objectives Vegetation health forecasting is an important field of research that involves predicting annual crop growth patterns in order to provide information to humanitarian aid agencies, government crop monitors, and famine warning systems. We propose and develop a novel algorithm to construct accurate 32-day-ahead image forecasts of vegetation health, as measured by NDVI (Normalized Difference Vegetation Index), over a 100 km by 100 km crop region in Ethiopia. Methods The vegetation health data was obtained from NASAs MODIS satellites. To improve our models ability to learn the annual patterns of crop growth, we developed a modification to the standard method of spatiotemporal forecasting, the ConvLSTM. Our proposed model, the Annual Gate ConvLSTM, introduces attention-network-weighted data from previous years into the memory cell of each recurrent unit. Furthermore, our variable length attention network improves upon the standard attention network by ensuring a constant standard deviation of weights for sequences of different lengths. This is achieved by multiplying all unnormalized terms by a value k, determined by a regula falsi approximation, before feeding the terms into the softmax normalization function. We measure the accuracy of the standard ConvLSTM and the Annual Gate ConvLSTM using Root Mean Square Error (RMSE, 0 is perfect, 1.3 is worst) on a common testing dataset of 70 sequences of 10,000 sq km NDVI data. Results Our Annual Gate ConvLSTM achieves RMSEs of 0.0852, 0.0834, 0.0882, and 0.0911 for the 8, 16, 24, and 32 day ahead predictions of NDVI values over the 10,000 square km region. The standard ConvLSTM achieves higher, less accurate RMSEs of 0.1005, 0.1135, 0.1218, and 0.1412 for the four respective timesteps. Furthermore, our network outperforms baseline models and existing methods of NDVI forecasting. Conclusions The superior performance achieved by the Annual Gate ConvLSTM in comparison to the standard ConvLSTM suggests that the usage of a variable length attention network in a Recurrent Neural Network helps to discover and utilize periodic trends in sequential datasets. Furthermore, the best existing algorithm for large scale NDVI forecasts only provides 8 day, 10 km by 10 km forecasts of 1 square km resolution at an RMSE of 0.09 - performing fewer steps ahead, smaller scale, and less accurate forecasts.
Summary statement
We develop a novel, period-aware modification to the ConvLSTM to construct accurate, long-term, large- scale image forecasts of vegetation health captured by NASAs MODIS satellites over a 10,000 square km region in Ethiopia.
Help received
We designed, developed, and tested our algorithm entirely on our own. However, we obtained our data from the repository of NASAs MODIS satellite imagery. Furthermore, we would like to thank Dr. Wassila Thiaw & Dr. Pablo Bocca for their advice pertaining to the importance of NDVI forecasting.
Competition history
- CSEF 2019
Resources
Related projects
CSEF · 2019
Predicting Optimal Farming Regions via Machine Learning Trained on Novel Vegetation Index
ISEF · 2026
Novel Synergization of Lunar Chronobiology, Genotypic Stress Analysis, and CNN-LSTM Architectures for Hyper-Localized Climate-Adaptive Agricultural Planning
ISEF · 2018
Predicting Food Shortages in Africa from Satellite Imagery
ISEF · 2023
Forecasting County Level Crop Yield in Top Agricultural-Producing States Using Satellite Data
AJAS · 2025
Creating a Drought Prediction Model Using Convolutional Neural Networks
JSHS · 2025
An Intelligent Bee Health Assessment System Using Cross-Attention-based Multimodal Neural Network for Visual and Audio Signals
ISEF · 2024
ForestCast: Developing a Novel Convolutional Neural Network to Enhance Prevention of Fragmentation in the United States
JSHS · 2024
Forecasting Post-Wildfire Vegetation Recovery in California using a Convolutional Long Short-Term Memory Tensor Regression Network
Closest projects by meaning, across every fair and year in the corpus.
Browse more like this
Source: California Science & Engineering Fair public projects