Improved Crop Classification with Rotation Knowledge using Sentinel-1 and -2 Time Series

Abstract

<jats:p>Leveraging the recent availability of accurate, frequent, and multimodal (radar and optical) Sentinel-1 and -2 acquisitions, this paper investigates the automation of land parcel identi- fication system (LPIS ) crop type classification. Our approach allows for the automatic integration of temporal knowledge, i.e., crop rotations using existing parcel-based land cover databases and multi-modal Sentinel-1 and -2 time series. The temporal evolution of crop types was modeled with a linear- chain conditional random field, trained with time series of multi-modal (radar and optical) satellite acquisitions and associated LPIS. Our model was tested on two study areas in France (<jats:italic>≥</jats:italic> 1250 km<jats:sup>2</jats:sup>) which show different crop types, various parcel sizes, and agricultural practices: . the Seine et Marne and the Alpes de Haute-Provence classified accordingly to a fine national 25-class nomenclature. We first trained a Random Forest classifier without temporal structure to achieve 89.0% overall accuracy in Seine et Marne (10 classes) and 73% in Alpes de Haute-Provence (14 classes). We then demonstrated experimentally that taking into account the temporal structure of crop rotation with our model resulted in an increase of 3% to +5% in accuracy. This increase was especially important (+12%) for classes which were poorly classified without using the temporal structure. A stark posi- tive impact was also demonstrated on permanent crops, while it was fairly limited or even detrimental for annual crops.</jats:p>

Mapping Crop Types of Germany by Combining Temporal Statistical Metrics of Sentinel-1 and Sentinel-2 Time Series with LPIS Data

“…Various studies have shown the potential to map crop types based on Sentinel-1 time series without additional optical data input in Europe (e.g., [30][31][32][33][34][35][36][37][38][39][40][41]) and in Germany (e.g., [42][43][44]). However, particularly promising are approaches combining optical and SAR data (e.g., for Asia [45], Africa [46,47], Europe [48][49][50][51][52][53][54][55], and Germany [56][57][58]). Such synergistic approaches consistently showed better overall accuracies than monosensor classifications (e.g., [48,51,56,57]), particularly in the case of challenging cloud situations [49].…”

Crop Rotation Modeling for Deep Learning-Based Parcel Classification from Satellite Time Series

“…Osman et al [31] propose to use probabilistic Markov models to predict the most probable crop type from the sequence of past cultivated crops of the previous 3 to 5 years. Giordano et al [32] and Bailly et al [33] propose to model the multi-year rotation with a second order chain-Conditional Random Field (CRF). Finally, Yaramasu et al [34] were the first to propose to analyze multi-year data with a deep convolutional-recurrent model.…”

“…Conditional Random Fields: M CRF . Based on the work of [32,33], we implement a simple chain-CRF probabilistic model. We use the prediction of the previous PSE+LTAE, calibrated with the method of Guo et al [40] to approximate the posterior probability p ∈ [0, 1] L of a parcel having the label k for year i:p k = P(l i = k | x i ) (see Section 3.3 for more details).…”

Crop Rotation Modeling for Deep Learning-Based Parcel Classification from Satellite Time Series

“…Osman et al [24] propose to use probabilistic Markov models to predict the most probable crop type from the sequence of past cultivated crops of the previous 3 to 5 years. Giordano et al [14] and Bailly et al [3] propose to model the multi-year rotation with a second order chain-Conditional Random Field (CRF). Finally, Yaramasu et al [38] were the first to propose to analyze multi-year data with a deep convolutional-recurrent model.…”

“…Conditional Random Fields: M CRF . Based on the work of [3] and [14], we implement a simple chain-CRF probabilistic model. We use the prediction of the previous PSE+LTAE, calibrated with the method of Guo et al [15] to approximate the posterior probability p ∈ [0, 1] L of a parcel having the label k for year i : p k = P (l i = k | x i ) (see Section 3.4 for more details).…”