A Hidden Markov Models Approach for Crop Classification: Linking Crop Phenology to Time Series of Multi-Sensor Remote Sensing Data

Abstract

Vegetation monitoring and mapping based on multi-temporal imagery has recently received much attention due to the plethora of medium-high spatial resolution satellites and the improved classification accuracies attained compared to uni-temporal approaches. Efficient image processing strategies are needed to exploit the phenological information present in temporal image sequences and to limit data redundancy and computational complexity. Within this framework, we implement the theory of Hidden Markov Models in crop classification, based on the time-series analysis of phenological states, inferred by a sequence of remote sensing observations. More specifically, we model the dynamics of vegetation over an agricultural area of Greece, characterized by spatio-temporal heterogeneity and small-sized fields, using RapidEye and Landsat ETM+ imagery. In addition, the classification performance of image sequences with variable spatial and temporal characteristics is evaluated and compared. The classification model considering one RapidEye and four pan-sharpened Landsat ETM+ images was found superior, resulting in a conditional kappa from 0.77 to 0.94 per class and an overall accuracy of 89.7%. The results highlight the potential of the method for operational crop mapping in Euro-Mediterranean areas and provide some hints for optimal image acquisition windows regarding major crop types in Greece. OPEN ACCESSRemote Sens. 2015, 7 3634

Multi-Temporal Land Cover Classification with Sequential Recurrent Encoders

“…In general, a direct numerical comparison of classification accuracies is difficult, since these are dependent on the number of evaluated samples, the extent of evaluated area and the number of classified categories. Nonetheless, we compare our method with the approaches of Siachalou et al [15] and Hao et al [13] in detail since their achieved classification accuracies are on a similar level as ours. Hao et al [13] used an RF classifier on phenological features, which were extracted from NDVI and NDWI time series of MODIS data.…”

“…In that regard, hidden Markov models (HMMs) [15] and conditional random fields (CRFs) [16] have shown promising classification accuracies with multi-temporal data. However, the underlying Markov property limits long-term learning capabilities, as Markov-based approaches assume that the present state only depends on the current input and one previous state.…”

Multi-Temporal Land Cover Classification With Long Short-Term Memory Neural Networks

“…Our LSTM model achieved good classification accuracies compared state-of-the art, while considering a notably larger number of crop classes (Foerster et al, 2012;Siachalou et al, 2015). While the hidden markov model approach of Siachalou et al (2015) is methodically closest to our deep learning strategy, their relatively small study area together with their small number of classes impede direct comparison.…”

“…While the hidden markov model approach of Siachalou et al (2015) is methodically closest to our deep learning strategy, their relatively small study area together with their small number of classes impede direct comparison. Our deep learning approach achieved better accuracy performance than the approach of Foerster et al (2012) using spectro-temporal NDVI profiles and adjusting these by additional agro-meteorological information (cf.…”

“…They use LANDSAT-7 images aggregated over several years from a large study area in north-east Germany and classify twelve crop classes in total. While these approaches follow a generic feature extraction and classification pipeline, Siachalou et al (2015) utilize a hidden markov model (HMM) approach which retains sequential consistency of multi-temporal observations on four LANDSAT-7 and one RAPIDEYE observation of Thessaloníki, Greece in 2010. Methodically similar to ours, Lyu et al (2016) LSTM Cell utilize RNNs and LSTM architectures to multispectral LANDSAT-7 and hyperspectral EO-1 HYPERION images, but-in contrast to our approach-for the purpose of change detection.…”