Publications

Bodani, P; Sharma, SA; Pandya, A (2025). Unified Multi-layer Perceptron Model for 16-Day MODIS NDVI Data Imputation and Short-Term Forecasting for Indian Region. JOURNAL OF THE INDIAN SOCIETY OF REMOTE SENSING.

Abstract
The pivotal role of the Normalized Difference Vegetation Index (NDVI) data in monitoring vegetation health and land surface dynamics is well acknowledged. However, cloud cover significantly impairs the accuracy of NDVI measurements, necessitating reliable preprocessing techniques, particularly for yield modeling and vegetation monitoring. This paper presents a novel multi-layer perceptron (MLP) model designed specifically for cloud imputation in NDVI time series data, utilizing the 16-day MODIS NDVI dataset for the Indian region. The primary contributions of our approach, compared to previous research, include (1) eliminating the need for synthetic filtering or extensive preprocessing by directly training on raw NDVI data, thus simplifying the workflow; (2) effectively leveraging existing quality flags in the MODIS dataset, enabling the model to inherently prioritize high-quality observations during training; and (3) developing a unified model capable of both interpolation (handling missing data across 1 to 4 consecutive time steps) and 1-step extrapolation tasks, thus providing enhanced flexibility and generalization for real-world data scenarios. Experimental results demonstrate that the proposed MLP-based model achieves improved accuracy, with a Mean Absolute Error of 0.035 for interpolation tasks and 0.044 for extrapolation tasks, outperforming an LSTM-based architecture particularly in extrapolation scenarios. Future directions will explore integrating attention mechanisms with spatial, climate, and weather contexts to further enhance prediction performance.

DOI:
10.1007/s12524-025-02193-0

ISSN:
0974-3006